1SYNTACTIC CONTROL OF TIMING IN SPEECH PRODUCTION
by
WILLIAM EDWIN COOPER
A.B., Brown University
1973
A.M., Brown University
1973
SUBMITTED IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE
DEGREE OF DOCTOR OF
PHILOSOPHY
at the
MASSACHUSETTS INSTITUTE OF'
TECHNOLOGY
February, 1976
Signature of Author.... rv..rrrn.r. rrs... .pur..............
Department of Psychology,
JanuaFy 12, 1976
Certified by .o/ffVKj/...' . ... j
Thesis Supervisor
Accepted by................. .......................
Chairman, Departmental Committee
on Graduate Students
MAI 4: 197)
19%RA lis
2SYNTACTIC CONTROL OF TIMING IN
SPEECH PRODUCTION
by
William Edwin Cooper
Submitted to the Department of Psychology in partial fulfillment
of the requirements for the degree of Doctor of Philosophy.
ABSTRACT
A sentence-reading procedure was used to study the influence
of syntactic and semantic structure on the timing of syllables in
speech production. In each sentence, easily segmentable key words
were placed in the environment of a putative syntactic boundary.
The immediate phonetic environments of the key words were held fixed
or independently assessed for their effect on syllable timing.
Chapter 1 includes an introduction placing the study within
a larger context of studies on complex motor skills that are segmented
and hierarchically organized. The chapter contains a discussion of
the study's principal motives, outcomes, and implications for theory
and practical applications.
Chapters 2 through 7 include descriptions of experimental
studies and their relation to issues discussed Chapter 1. The
experiments of Chapter 2 showed that speakers shortened the durations
of segments just prior to the locations of two syntactic deletions
which erased material from the beginning of a clause, whereas speakers
lengthened the durations of segments just prior to the location of a
deletion erasing material from the middle of a clause. The results
provided support for the notion that a syntactic level of processing
controls timing and that this level computes a syntactic representation
that corresponds more closely to linguistic surface structure than
underlying structure.
Chapter 3 includes a study of speech timing at the clause
boundaries of complement structures. Two types of complement were
studied, triggered in the main experiment by the verbs expect and persuade.
The results provided further support that a syntactic level of processing
controls speech timing and that this level computes a clausal representa-
tion similar to linguistic surface structure. An auxiliary experiment
showed that a phonetic influence on syllable timingconditioned by the
presence of a following voiced vs. voiceless segment, extends across a
verb-noun phrase boundary.
3Chapter 4 contains a report of experiments on the possible
effects of syntactic rules of preposing on timing. The rules
Passive, Topicalization, Adverb Preposing, and Prepositional Phrase
Proposing were considered. The results provided no support for the
possibility that morphologically null traces, as defined in current
linguistic theory, are accompanied in speech production by changes
in the timing of immediately surrounding words.
Chapters 5 and 6 include experiments showing that the semantic
relation of coreference plays a role in the control of timing. In
the experiment reported in Chapter 5, speakers lengthened the duration
of a noun when it was referred to again later in the same sentence by
a pronoun. This effect was observed for the very first word of an
utterance and for non-adjacent coreferents that spanned a major clause
boundary. A second experiment, reported in Chapter 6, added further
support to the notion that coreference relations control timing and
that such control is not restricted to the domain of a single clause.
Chapter 7 includes a sketch of additional experiments for which
a relatively small data base was obtained. Three major issues in timing
research are discussed in conclusion with reference to the present study
and plans for further work.
Thesis Supervisor: Merrill F. Garrett
Title: Associate Professor of Psychology
4TABLE OF CONTENTS
Page
Title............................-..--------------------.1
Abstract........................................---------2
Table of Contents......................................4
Acknowledgements...............................----5
Biographical Sketch...................................--6
Note to the Reader............................... --..7
Chapter 1. Introduction......................-..----..8
Figure 1..............................----.--.----46
Figure 2.......................................----47
Chapter 2. Syntactic Control of Segment Length in
Speech..................................................-48
Chapter 3. Syntactic Control of Speech Timing: A
Study of Complement Clauses..................-.........58
Figure 1....................--.-....-------------109
Chapter 4. Preposing Rules...........................110
Chapter 5. Speech Timing of Coreference..............129
Chapter 6. Speech Timing of Coreference. II.
Precede and Command Relations.........................136
Figure 1...................-.-------------------147
Chapter 7. Conclusion.............................-.148
5Acknowledgements
I wish to thank the following individuals who have been
particularly helpful to me during the course of this thesis:
Dumont Billings, Irene Bonner, Noam Chomsky, Ann Cutler, Thomas
Felton, Jerry Fodor, Merrill Garrett, John Groopman, Richard Held,
Dennis Klatt, Steven Lapointe, Cornelia Parkes, John Robert Ross,
Kenneth Stevens, Hans-Lukas Teuber, Luis Tiant, and Robin Tuckerman.
[see individual chapters for specific acknowledgements].
6Biographical Note
The author was born on March 20, 1951 in Baltimore, Maryland.
He attended public schools in Lancaster, Pennsylvania from 1956 to
1969. During the 9th grade, he received local recognition for placing
three root beer Fizzies in the water bucket of a certain algebra
teacher, just prior to the teacher's routine washing of the blackboard
(thereby ending a promising career in mathematics). From 1969 to
1973, he attended Brown University. From 1973 to 1975, he attended
graduate school at M.I.T., where he divided his TV viewing time
between Star Trek and You Bet Your Life. He has published numerous
articles, one of which (co-authored by Professor J. R. Ross) was
discussed in a recent column of the Christian Science Monitor. His
professional ambition, presently elusive, is to learn to do science
with pride and without arrogance.
7Note to the Reader
Some of the chapters in this thesis were written as self-
contained papers, intended for journal publication. For this
reason, the format differs slightly from chapter to chapter, and
there exist a few instances of unnecessary repetition. All references
appear at the ends of individual chapters.
8Chapter 1
Introduction
9The aim of this chapter is to provide an overview of the study
by describing its motivation, outcome, and applications. The chapter
is intended for anyone who might be curious about how people speak,
as well as for some who, by their own admission, are not at all curious
about this topic, but who are interested nonetheless in the more general
problem of information-processing by organisms.
Section 1 includes a statement of a major assumption of the study,
as well as some discussion of the possibility that this assumption
represents a general property of information-processing in motor skills
which exhibit two particular properties -- segmentation and hierarchical
organization. Section 2 contains remarks on previous theory and research
on speech production insofar as such work has guided the present venture.
Section 3 includes a description of the study's principal aims, methods,
and results. Finally, Section 4 includes an outline of a revised
theory of speech production and the implications of the study for
further research and health- and engineering-related applications. At
worst, a look at this chapter should allow the reader to make a
relatively informed decision about whether his interests would be
served better by reading the study in detail or by turning attention
instead to a nearby journal, game of squash, or any of a number of other
attractions.
10
1. Aside from general assumptions concerning the orderliness of
nature, much of this study was based on one major assumption about a
relation between centralized motor commands and motor output. In its
general form, the assumption can be stated as follows:
(1) Given a domain of information processed by the central nervous
system with an ordered sequence of segmented contents
s ,S2' s3...sn 3 , the motor output controlled by fsn 3
will be longer in duration than the average of the segments
controlled by the processing domain as a whole, ceteris paribus.
We may refer to this assumption as domain-final or phrase-final lengthening.
The assumption is simple, intuitively appealing, and, as we shall see,
can be used to pursue an understanding of processing domains in speech
production, an area about which very little is currently known.
The intuitive appeal of the assumption can be pinpointed for
certain types of behavior, including speech. If a listener attends to the
durations of speech sounds during a conversation, he notices that
syllables which occur at the ends of sentences and at the ends of
certain clauses are lengthened in comparison with their typical durations.
In addition, the lengthened segments are often followed by a pause.
These impressions of the unaided ear have been confirmed by actual
measurements of acoustic duration (Martin, 1970; Lindblom and Rapp, 1973;
Klatt, 1975a; Kloker, 1975).
11
Before considering speech in detail, however, it should be noted
that the appeal of the phrase-final lengthening assumption extends to
other forms of behavior. A good example is musical composition, for
which, like speech, we have some idea on independent grounds about the
boundaries of domains of processing, including movements, stanzas,
phrases, and the like.1 The independent grounds are provided by the
theory of musical composition as by the theory of grammar. The unaided
ear tells us that musical notes are generally longer at the ends of
phrases than the average duration of notes within such phrases. This
lengthening effect occurs with remarkable regularity in classical and
modern music of the Western world with which I am familiar. The
reader may wish to test this claim by listening to the radio for a
half-hour or so. Although a number of other factors may influence the
durations of individual musical notes (see below), the phrase-final
lengthening effect appears consistently throughout most compositions.
One of the more famous examples of the principle is taken from
the score of Beethoven's 5th Symphony, consisting of a phrase of four
notes -- three short followed by one long. If the ordering of the
short and long notes were reversed, the phrase would sound not only
unfamiliar, but, I would argue, unnatural as well. According to this
view, a musical phrase beginning with a long note and ending with a
short one is an unnatural output for a composer, other factors being
equal.
12
The general principle of phrase-final lengthening appears to be
ingrained in the listener as well as the composer. A simple experiment
in auditory perception confirms this notion. If a listener is presented
a tape loop containing two tones of the same frequency and amplitude,
but of different durations, the listener hears the sequence of tones
as a pair of short-long tones, not as a pair of long-short tones (Allen,
1975). It seems, then, that the auditory system naturally processes
information in terms of a domain containing short-long rather than long-
short sequences, paralleling the constraint on the motor output of
both musical composer and speaker.2
Two further points should be made about the nature of phrase-final
lengthening for speech and music. First, it seems that the general
principle stated in (1) might be represented more precisely by taking
into account different magnitudes of lengthening for processing domains
of different hierarchical status. In both speech and music, there
exists a general trend for the lengthening at the ends of minor phrases
to be less pronounced than the lengthening at the ends of successively
more inclusive domains. Thus, the end of a multi-clause sentence
is generally longer than the end of an individual clause within that
sentence, which is in turn longer than the end of an individual phrase
within that clause, according to impressions of the unaided ear.
Similarly, in a symphony, the end of a major movement is generally
longer than the end of a lesser domain.
13
A second point concerning speech and music, already hinted at,
is that the principle in (1) should hold with absolute certainty only
when other factors that might influence segment durations are held
equal, which they of course seldom are. Such factors may include
changes in overall rate of movement, fatigue,3 trade off between
4
amplitude and duration, as well as factors that rely on relatively
complex constraints on information-processing. Unfortunately, too little
is known about most of these factors to provide a flawless, independently
motivated account of any cases of segment length which violate the
general principle proposed above. Nevertheless, the available impressions
of the listener for both speech and music are sufficiently supportive
of the principle that it seems justifiable to claim that the principle
is not merely intuitively appealing but valid, at least within the
realms of speech and music.
Speech and music provide the best available sources of information
about the general principle because in these cases there exist independent
grounds for defining domains of processing, based on the theories of
grammar and musical composition. Since the domains can be independently
defined, any possible circularity in testing the principle can be avoided.
For other forms of complex motor behavior involving segmented information,
there typically exist weaker, sometimes meager, grounds for defining
the bounds of processing domains. However, for the few cases where such
14
grounds are relatively well-established, the evidence provides support
for the principle of phrase-final lengthening. We turn now to consider
these cases in an attempt to determine the necessary and sufficient
conditions under which the general principle applies.
The communication systems of other species provide a few tests
of the general principle, and in most cases where I have been able to
find a documented difference in lengthening among phrase segments,
5.
the differences are in support of the general principle. In acoustic
traces of the song phrases of the chaffinch (fringilla coelebs), for
example, phrase-final segments are typically longer than segments at
the beginning or middle of a phrase (see Figure 1). This claim is
based on an examination of the acoustic traces published by Thorpe
(1961) and Nottebohm (1970). Unfortunately for present purposes,
however, these and other researchers of birdsong have been generally
concerned with analyzing acoustic traces in terms of the frequency domain
rather than in terms of duration, and to my knowledge no highly systematic
Insert Figure 1 about here
treatment of segment durations in birdsong has yet been published.
Tf the spectrographic traces published by Nottebohm (1970) are
representative of the durational characteristics of chaffinch song, a
second major point about birdsong duration is worth noting. Phrase-final
15
lengthening is found not only in traces of the adult song but in the
earliest stage of song development as well, even for birds reared in
the absence of an auditory model. If the latter condition applies
to phrase-final lengthening generally, it would provide evidence that
the lengthening effect is genetically pre-determined.
A better documented example of phrase-final lengthening in
animal communication involves chirps emitted by the insect Ambyorypha
oblongifolia, a member of the family Orthoptera (du Mortier, 1963, p. 359).
This insect produces chirp phrases consisting of four segments, having
durations averaging 8, 13, 21, and 38 msec, successively. The example
illustrates another possible refinement of the general principle in (1),
namely that gradations in length occur within a processing domain,
making the lengthening effect observed in absolute phrase-final position
a special case of a duration principle that applies to each segment
of a phrase, other factors being equal.
A consideration of acoustic communication in other species has
suggested that phrase-final lengthening is not restricted to our own
species. In addition, more systematic study of animal communication may
confirm the possibility, suggested by an examination of published
birdsong traces, that phrase-final lengthening is a genetically pre-
determined characteristic of acoustic communication.
16
In addition to birdsong and insect chirps, one might expect to
find tests of the general lengthening principle in other types of animal
communication, in particular, those that have been studied in some
acoustic detail, including the sound emissions of the dolphin (Lilly,
1963, 1965), squirrel monkey (Winter, Ploog, and Latta, 1966), bullfrog
(Capranica, 1965, 1968), and others (see Busnell, 1963, for a review).
Yet, as in the case of birdsong, research on these species has not
treated segment timing systematically, but has been primarily concerned
with the frequency domain. The results of this study on speech timing
suggest, however, that a similar research strategy might be applied with
profit in other areas of animal communication, where, as in the case
of speech, very little is known about the hierarchical structure of
processing.
In search of other properties of the general principle, we can
consider non-acoustic forms of motor output as well as non-communicative
outputs. In our species, non-acoustic forms of communicative behavior
include handwriting, telegraphy, sign language, and dance, all of which
are segmentable to some extent. My intuition suggests that phrase-final
lengthening exists for these behaviors, independent of other factors
that may influence movement duration (such as the need for precision
in reaching a goal). However, I have been able to find no documented
evidence either in support of9 or in opposition to, the phrase-final
lengthening principle for these behaviors. In other animals, relevant
17
activities include communicative insect dancing (von Frisch, 1950)
and perhaps certain non-communicative behaviors as well, including
pecking in chicks and swimming in fish, to the extent that such behaviors
can be shown on independent grounds to be hierarchically organized and
segmentable. So far, all evidence in support of phrase-final lengthening
comes from communicative behavior, and it remains a moot question
whether communicative function is a necessary condition for the operation
of the principle.
The phrase-final lengthening principle is used in this study
primarily as a starting point for probing the control of timing in
speech, and we need not know why the principle holds to undertake
this effort. Yet, the question will ultimately become an important one
for speech theorists, and we will consider here three possible answers
that apply generally to the various behaviors that exhibit such
lengthening.
Two types of explanation can be classified as transmitter-
oriented, while a third can be classified as receiver-oriented. One
of the transmitter-oriented accounts states that organisms produce
phrase-final lengthening in order to permit an extra fraction of timing
during which to compose a following phrase. This account rests on the
reasonable assumptions that (a) motor output is planned on a phrase-by-
phrase basis to some extent and (b) the planning of an upcoming phrase
occurs primarily near the end of the currently produced phrase. This
18
account may be naturally extended to provide an explanation of the
high frequency of pauses occurring between phrases, on the assumption
that planning continues during the pause interval. The planning
account, however, cannot explain another related finding, namely that
segments are usually lengthened at the ends of discourses, symphonies,
etc., where no further planning is required. It is conceivable, of
course, that the planning account provides a correct explanation of
phrase-final lengthening but that an independent principle of timing
accounts for the lengthening observed at the termination of discourses
and musical compositions. The latter principle might take the form
of a receiver-oriented principle, which operates under the assumption
that exaggerated lengthening at the end of the entire behavior pattern
serves as a cue to the listener that the behavior has in fact ended.
The second transmitter-oriented account of phrase-final lengthening,
like the planning account, relies on the assumption that the timing
of a phrase is planned in a unitary fashion at some processing level.
According to this second account, however, lengthening is produced
at the ends of phrases not to aid the planning of upcoming phrases but
as a consequence of the manner in which the previously planned unit was
timed and stored before segment output. For example, segments of a
processing domain might be stored in a buffer which operates like a
push-down store (Simon and Kotovsky, 1963)6, containing a "spring"-like
mechanism. The force on the spring is directly proportional to the
19
number of segments currently in storage. As a consequence, successive
segments are emitted from the buffer at successively slower rates,
and hence, possess successively longer durations. This account provides an
explanation of the graded lengthening effect noted for insect chirps,
although the planning account can also be modified to handle this case.
Discourse-final lengthening is not handled well by either the planning
or buffer accounts, and there is generally little or no a priori reason
to favor one of these accounts over the other.
A third general account of phrase-final lengthening is based
on the assumption that lengthening is produced as a cue for the receiver,
a cue which aids the receiver in recovering structural properties of
the message. Unlike the transmitter-oriented accounts, the receiver-
oriented explanation is limited to communicative behavior, but, as we
have noted above, all known instances of phrase-final lengthening belong
to this category.
The three general types of explanation for phrase-final lengthening
are not incompatible with one another. This fact makes it possible that
more than one of these accounts is responsible for the lengthening
effect observed in any particular case, increasing the degrees of
theoretical freedom and making it very difficult to provide tests that
distinguish among the three main alternatives. To complicate matters
further, it should be pointed out that, in addition to the three
general alternatives, there exist certain specific accounts which may
20
apply to particular forms of phrase-final lengthening. An example
of such an account occurs with speech, where it has been suggested
that phrase-final lengthening is produced to permit a longer interval
during which to produce intonational changes in phrase-final position
(see Klatt, 1975a). Another possibility for speech, suggested by
Stevens (personal communication), is that phrase-final lengthening is
produced to allow the speaker time to reset laryngeal and articulatory
postures for the upcoming phrase. This view is based on the notion
that such postures "run down" during the production of a phrase and
thus require resetting. Work has not proceeded to the point where this
hypothesis can be formulated in precise terms, however. Eventually,
it should be possible to provide critical tests of the alternative
accounts noted here, but a number of other facts need to be tacked
down beforehand. The present results are accountable in terms of any
of the explanations cited above, with one important class of exceptions.
Some of the consistently observed effects on duration were very small
in average magnitude (^10 msec), too small to be reliably detected by
listeners. Accordingly, such effects cannot be accounted for in
terms of the receiver-oriented hypothesis noted above.
21
2. We turn now to consider speech timing in some detail . Of the
various behaviors discussed above, speech is believed to be the best
area in which to undertake an in-depth study of timing control.
The reason is that speech is our most vital communicative activity,
and the study of timing control in this area should allow us to apply
proper constraints on a theory of speech production, a matter of both
intellectual and practical value. On one hand, providing a refined
theory of speech production would bring us closer to understanding the
essence of mental operations. On the other hand, such a theory can
be applied directly to problems in communications engineering and
speech pathology. These considerations, discussed more concretely
below, provided the general motivation for the present study.
It was noted at the outset of this chapter that speech segments
tend to be longer at the ends of phrases, clauses, and sentences
than in non-final positions. For some types of sentence structures,
this lengthening effect is clearly audible. It is not very surprising,
then, that acoustic analyses of speech signals have confirmed this
impression. And yet, such studies have brought to our attention in
a particularly direct manner the fact that the control of speech
timing is a rather intricate process, involving not only relatively
low-level properties of the articulators, but syntactic properties
of the utterance as well.
22
Martin (1970) discovered that grammatical boundaries were typically
accompanied by lengthening of the preceding syllable in a spectrographic
study of English spontneous speech. The speech corpus was not very
large (only 60 utterances) and the measurement technique was relatively
crude, yet Martin's results did show the presence of a general clause-
final lengthening effect. Kloker (1975) has corroborated and extended
this finding for spontaneous speech, and Lindblom and Rapp (1973) and
Klatt (1975a) have found similar effects for practiced reading, using
Swedish and English, respectively. The combined results of the studies
lend strong support to the notion that the ends of major grammatical
domains are accompanied by segment lengthening.
Each of these studies contained two limitations which are
important from the standpoint of the present study. First, with
the exception of one aspect of the study conducted by Lindblom and Rapp
(1973), none of the studies involved tests of syntactic effects that
controlled for other influences known to affect lengthening, in particular
influences of sound structure and sound environment (see Lehiste, 1970,
for a review of some of these factors). Klatt (1975a) took such factors
into account in a post hoc analysis, but experiments were not conducted
to test comparisons between phonetically-matched sentences differing
only in syntactic variables. Second, none of the previous studies included
direct tests of a variety of syntactic boundaries, or cases in which the
23
locations of syntactic boundaries were controversial according to
competing linguistic theories or according to distinct levels of
representation within a given theory.
The present study was designed to provide a high degree of
phonetic control and at the same time provide a testing ground for a
number of hypotheses about the control of speech timing by syntactic
(and semantic) factors. The information available about the influence
of other factors on timing control is fairly extensive, enabling us
to control for such factors wherever possible or to take their influence
into account when their presence must be tolerated to test a syntactic
hypothesis.
The additional influences on speech timing include gross factors
such as overall speaking rate (Malecot, 1969; Gay, Ushijima, Hirose,
and Cooper, 1974), word emphasis (Lieberman, 1967; Bolinger, 1972),
word frequency (Coker, Umeda, and Browman, 1973), as well as detailed
properties of sound structure, including the inherent duration of
segments (Klatt, 1975a) and effects of immediate phonetic environment.
An example of the last type of effect i the influence of a following
voiced vs. voiceless consonant (e.g. [b] vs. [p]) on the duration of an
immediately preceding vowel -- the vowel duration is longer when
followed by a voiced consonant (Peterson and Lehiste, 1960; House,
1961; Delattre, 1966; Lisker, 1974). And, in addition to such
24
structural effects, the overall phonological stress pattern (Fry,
1955; Chomsky and Halle, 1968) plays a major role in the control of
timing.
In normal speech, the interplay of these various factors may
obliterate the audibility of phrase-final lengthening, although not
to the extent that such lengthening is not generally detectable during
the course of a conversation. From a practical standpoint, the presence
of such a great number of factors necessitates that a detailed study
of syntactic control of timing neutralize these other factors wherever
possible. From a theoretical standpoint, the presence of so many
extra factors suggests that either (a) speech timing and its possible
perceptual relevance make use of extremely complex processing machinery,
or (b) a much smaller number of control factors is accountable for the
wide range of influences observed in speech behavior.
A version of a currently popular model of speech production is
presented in Figure 2. The model represents a slightly more detailed
Insert Figure 2 about here
version of one described by Liberman (1970), among others. The
principal features of the model are its strictly serial stages of
processing and the close correspondence between such stages and components
of a generative grammar (Chomsky, 1965). The model is presented here
25
primarily because it provides a useful framework for introducing
aspects of the present study, not because it holds any special claim
to validity.
The first stage of the model represents the conversion of thought
into some form of linguistic representation. It is typically assumed
that the linguistic representation is in essence a semantic representation,
although it need not be (Fodor, 1975). Next, an underlying syntactic
structure is formed, analogous to the level of deep structure proposed
in generative grammar (see Postal, 1964 for introductory motivation
for the linguistic distinction between deep and surface structures).
Lexical items are inserted after the underlying structure has been
formed, resulting in an output roughly corresponding to a terminal
string in the grammar of Chomsky (1965). A system of transformations
operates to move, add, or delete elements from the structure. The
output of the transformational stage is a syntactic surface structure.
This structure is transmitted to the phonological component of the
speech processor, where word- and syllable-level rules are applied
to produce the phonetic output, including the desired phonetic structure
as well as proper stress and timing relations.
If the general outline of this serial model is correct, then
the control of timing relations in the phonological stages of processing
should be influenced by syntactic and/or semantic information only to
the extent that such information is preserved at the final stage of
syntactic processing. That is, information available in the surface
structure representation should be capable of influencing the control
26
of timing, but not information about semantic structure, underlying
syntactic structure, or transformational derivation, which does not
also appear in surface structure.
The model is speculative and rests on no more than a few strands
of experimental evidence. The prediction of the model concerning
timing control has not been tested, nor have any of the previous
studies on phrase-final lengthening provided information that has
d3I~rte-artIng--on the prediction. The previous evidence which does
concern the general form of the model involves analyses of errors in
spontaneous speech (Fromkin, 1971; MacKay, 1972, 1973;
Shattuck, 1974; Garrett, 1975), a few studies of hesitation phenomena,
as well as some experimental attacks on the problem (see Fodor, Bever,
and Garrett, 1974, Chapter 7, for review). The data from speech
errors suggest that more than a single syntactic level of processing
may exist in speech production, as required by the present model
(Garrett, 1975). However, error analyses and other tests have so
far failed to uncover any detailed properties of levels of syntactic
processing other than properties of a level that corresponds to a
surface structure representation.
A major drawback facing someone interested in discovering
properties of the speech production system is the lack of an experimental
paradigm that can be applied validly and efficiently to a large
sub-set of the problems at hand. Those who analyze speech errors have
27
properly laid claim to the face validity of their enterprise, but while
the analysis of errors has provided a very useful source of information,
the work involved in obtaining a sufficiently large number of errors
relevant to testing a particular hypothesis is at best arduous and,
in some cases, impossible.
In addition to the general model of speech production reviewed
above, a number of models have been proposed to account for speech
timing. One model relies on the notion of isochrony, which states
that speakers attempt to produce the onsets of stressed syllables
at approximately equal intervals (Kozhevnikov and Chistovich, 1965;
Lehiste, 1970). Both methodological and theoretical objections have
been raised against the isochrony principle (Ohala, 1970; Klatt, 1975a).
A fairly elaborate theory of stress timing, based in part on the isochrony
principle, has been advanced by Martin (1972) to account for timing
relations in both speech and musical composition. Other models of
timing have been concerned primarily with two major processing distinctions,
the difference between linear vs. hierarchical planning (Lashley,
1951) and between central vs. peripheral feedback control (MacNeilage,
1970; Ohala, 1970). These distinctions will be of marginal use in
providing an account of the results of this study.
28
3. The principal aims of this study were:
(1) to determine whether surface, transformational, and/or
underlying syntactic stages of processing control speech
timing
(2) to determine whether semantic relations control speech
timing
(3) to determine the domain of processing at any stage of
processing for which timing control can be shown.
In order to provide adequate tests of these questions, the control
of timing was studied for a number of different syntactic structures
with a single experimental design. The basic intent of the design was
to provide a means of assessing the timing effects of syntactic variables
while holding phonetic variables fixed as much as possible. To achieve
this goal, a sentence-reading task was employed in which speakers were
asked to read sentences as if they were uttering the sentences spontaneously.
Ideally, analyses of spontaneous speech itself would provide the most
fitting tests of a speech production model, but such analyses are strictly
impossible in cases like the present where tight control over phonetic
and situational variables is required. The sentence-reading task allows
the experimenter to control for these variables, and yet provides a
relatively natural speech situation. Since previous data on phrase-
final lengthening showed similar effects for both spontaneous speech
29
and practiced reading (Martin, 1970; Lindblom and Rapp, 1973; Klatt,
1975a; Kloker, 1975), there is some reason to believe that the syntactic
effects observed in practiced reading will be applicable to spontaneous
speech.
Each hypothesis in this study was tested using one or more lists
of sentences. Each list contained from 2 to 39 sentences, and each
sentence in a list contained one or more key words. The key words, to
be measured for duration, were placed at locations just before or after
a putative syntactic boundary (a different procedure was used in tests
of semantic variables; see Chapter 5). .Wherever possible, the key words
appeared in the middle of each sentence string, in order to minimize any
possible effects of changes in subglottal pressure on timing (Lindblom
and Rapp, 1973). In addition, the key words were selected on the basis
of their phonetic structure, so that each word was readily segmentable
in the speech waveform.
The sentences of a given list were closely matched for total number
of words and syllables. In addition, the lexical material and sentence
stress contour of the sentences were matched wherever possible. Finally,
8
the sentences of a list were equally plausible, to a first approximation .
Speakers were tested individually in a sound-insulated room. A
typical experimental session lasted about 45 minutes. During this time,
the speaker was given from 5 to 7 sentence lists. At the outset of the
session, the speaker was told that the general purpose of the experiments
30
was to study the control of speech timing in a relatively natural speech
setting. The speaker was then provided general instructions that concerned
the format of all tests in the session. These instructions included the
following main points: (a) a number of different sentence lists will be
presented; each list should be treated as a separate experiment, and each
sentence within a list should be treated independently of any other
sentences; (b) when a sentence list is presented, the speaker should
first practice saying the sentence until he is comfortable with the
utterance and is satisfied that he is able to utter the sentence with
normal rhythm as a unitary whole, not word-by-word as in unpracticed
reading; (c) emphatic or contrastive stress should not be placed on any
words or syllables in a sentence (this instruction did not apply to the
semantic test in Chapter 5); (d) after the speaker finishes practice,
he should inform the experimenter and then utter one more practice
utterance to allow the experimenter to check for undesirable emphatic or
contrastive stress and to check recording levels; (e) on cue from the
experimenter, the speaker should then say the sentence 6 times (or 2 times,
as noted for particular experiments) in succession; (f) if the speaker
departs from his normal practiced rhythm or utters a mispronunciation
during recording, he should utter the word "repeat", pause, and then say
the utterance again as necessary until the appropriate number of correct
occurrences of the sentence is obtained. For most tests, the speaker
31
practiced and read a given sentence for recording before practicing any
other sentences in the list; in some tests, as noted in later chapters,
the speaker instead practiced each sentence in the list before reading any
of the sentences for recording purposes.
In addition to these general instructions, each speaker was given
particular instructions for certain lists, as indicated in later chapters.
After the completion of each list, the speaker was usually encouraged
to take a drink of water, and a longer rest period was provided about
halfway through the test session.
For tests in which 6 occurrences of each sentence were recorded,
the first 5 occurrences were digitized at a sampling rate of 10 kHz and
analyzed for duration with the aid of a computer controlled cursor (Huggins,
1969). The reliability of the duration measurements varied slightly from
experiment to experiment, as indicated later, depending on the phonetic
structure and environment of the key word or segment. In all cases,
however, the reliability was estimated to be within + 5 msec and in most
cases within +3 msec (see individual chapters for reliability estimates).
The sentence-reading technique outlined here provided a fairly
efficient means of testing a variety of hypotheses about the organizational
structure of speech timing control. Further efforts to automate
aspects of the measurement procedure should make the technique more
desirable as a tool for studying sentence production.
32
We now turn to consider the main findings of the study. These
can be summarized as follows:
(1) evidence indicates rather strongly that a surface level
of syntactic structure controls speech timing; preliminary
evidence indicates that transformational and underlying
syntactic structure may also control timing
(2) the magnitude of clause-final lengthening differs as a
function of the particular type of clause
(3) semantic relations of coreference control timing
(4) the domain of semantic processing which controls timing
is not restricted to a single clause; however, at least
one domain of syntactic processing which controls timing
probably is so restricted
A number of specific findings were obtained in the study, but they
will not be reviewed here because they rely on linguistic constructions
whose properties have not yet been discussed.
4. The principal findings of the study indicate the need for important
revision and extension of the theoretical model outlines in Figure 2.
In particular, the model must now be re-constructed so that higher
level semantic information plays a role in the control of timing, either
by a direct route or by some other route not presently available in
the model.
33
In addition, the results suggest that semantic and syntactic
control operate over different domains. Semantic processing is not
limited to the domain of a single clause, whereas certain syntactic processing
does appear to be so limited. Currently, work is being directed at
specifying other details of the model (see Cooper, in preparation).
It was noted earlier in this chapter that one of the prime
motives for undertaking this study was the expectation that it would
be useful in guiding work in related areas of research. The outcome
of the research has confirmed this expectation, and we note below three
specific areas in which the results may play a guiding role.
One area, quite directly related to the work on speech production,
involves studies on the perception of duration. In a recent report,
Klatt and Cooper (1975) showed that listeners could detect differences
in the durations of speech sounds of the same order of magnitude as
the differences in duration produced by speakers, as a function of
the syntactic environment. The results of the current study and
further work using the sentence-reading paradigm provide important
information about the magnitudes of speech production differences,
information that can be used to guide the perceptual work. For example,
cases can be distinguished from the production data in which lengthening
effects are either clearly too small and inconsistent to be of perceptual
relevance or are well within the range of a listener's detectability.
34
Perceptual research can thus be directed at studying the latter cases,
in order to find out whether listeners can not only detect but utilize
duration information to recover structural properties during sentence
perception.
In addition to guiding work on sentence perception, the present
results provide useful input to programs designed to synthesize speech
from a phonetic transcription (Coker, Umeda and Browman, 1973; Klatt, 1975b).
Currently, such programs include durational rules to some extent in
order to produce more natural-sounding speech, but these rules typically
do not involve syntactic or semantic variables or else treat such
variables at only a very general level. In Klatt's (1975b) program,
for example, a clause-final lengthening rule is included to increase the
duration of clause-final segments by a small amount. However, the
results of the present study indicate that the magnitude of the lengthening
effect observed for clauses of different types may differ substantially.
These results can be used to implement a more refined set of clause-
final lengthening rules which take these differences into account.
The results of similar studies should permit the implementation of
sufficiently precise lengthening rules whose combined effect will be
to noticeably enhance the naturalness and understandability of synthetic
speech.
35
Improvements in speech synthesis have social significance
because synthesis programs are implemented as a component of reading
machines for the blind (Allen, 1973) and as part of an overall effort
to achieve full man-machine communication by speech. The latter
effort represents a much-desired but very remote goal which, if attained,
would revolutionize information-processing as we know it in daily life.
Typing would become obsolete, and many interactions within and among
businesses, education, and government would take place over telephone
lines between man and computer. Although this goal is well beyond
the reach of current understanding and technology (see Reddy, 1974),
results such as those provided here should play a small part in directing
the effort.
A third application of the present study concerns the teaching
of speech rhythm in the deaf. Boothroyd, Nickerson, and Stevens (1975)
have noted that improper rhythm is one of the more important drawbacks
in the speech of congenitally deaf children, and teachers of these
children are often unaware of the importance of rhythmic structure in
speech training. Work on phrase-final lengthening and other aspects
of speech timing may culminate in a system of durational rules that
could be taught to deaf children. However, like the engineering goal
of man-machine intercommunication by speech, this application can only
be attained after a much greater portion of the systematic research
effort on speech timing is completed.
36
Footnotes
1. Note that the theories of grammar and musical composition provide
an independent means of determining domains of structure but
not domains of processing in a psychological sense. We adopt
as a working assumption the principle that an isomorphic relation
exists between domains of structure and domains of processing
at some stage of central motor activity.
2. It seems that the poet should also be included in this group,
since iambic meter is more common than trochaic. Of related
interest is the observation by Allen (1975) that languages
having accent based primarily on duration (e.g. French) have
stress on final syllables of words, whereas languages having
strong tonic accent (e.g. English, German) have syllable-initial
stress primarily.
3. In most cases, general fatigue can be discounted as an explanation
for lengthening by showing that a segment at the end of a
phrase is longer than the average duration of segments in an
immediately following phrase.
4. It is expected that in cases where the phrase-final segment is
short, it will be of high amplitude. This effect is observed
in some modern symphonies.
37
5. The only documented counter-example that I have found so far
concerns the bird song of Pileated Tinamou, cited by Thorpe
(1961: 3). The song phrase of this species consists of note
groupings that possess rising gradations in pitch and gradations
of decreasing length.
6. The assumption that the buffer operates like a push-down store
can be applied to this situation only if another, counterintuitive,
assumption is made, namely that speech segments of a domain
enter the store such that the domain-final segment enters the
store first.
7. In addition to phrase-final lengthening, there exists another
speech phenomenon which supports the general prindiple in (1).
Cooper and Ross (1975) have noted that, in pairs of conjoined words
for wich the linear ordering of the words is rigidly fixed
(e.g. kit and caboodle/ *caboodle and kit) , words which are fixed
in second position have a vowel with an inherently longer duration
than the vowel of the word fixed in first position (e.g. stress and
strain, hem and haw), other factors being equal or taken into
account. The vowel length principle for fixed conjuncts
suggests an instance in which the principle in (1) has become frozen
in the structure of the language itself, in addition to operating
38
to influence the lengths of speech segments during on-line
speech production.
8. Since the reading of test sentences is practiced prior to
recording, it is anticipated that small differences in plausibility
will play a negligible role in determining segment length.
39
References
Allen, G. D. (1975). Speech rhythm: its relation to performance
universals and articulatory timing. Journal of Phonetics 3, 75-86.
Allen, J. (1973). Speech synthesis from unrestricted text. In J.L.
Flanagan and L. R. Rabiner (Eds.) Speech Synthesis. Stroudsburg,
Pa.: Dowden, Hutchinson, and Ross, Inc.
Bolinger, D. (1972). Accent is predictable (if you'te a mind-reader).
Language 48, 633-644.
Boothroyd, A., Nickerson, R.S., and Stevens, K.N. (1975). Temporal
patterns in the speech of the deaf: a study in remedial training.
Submitted for publication.
Busnel, R.-G. (1963). Acoustic Behavior of Animals. Amsterdam:
Elsevier.
Capranica, R.R. (1965). The Evoked Vocal Response of the Bullfrog.
Cambridge: M.I.T. Press.
Chomsky, N. (1965). Aspects of the Theory of Syntax. Cambridge: M.I.T. Press.
Chomsky, N. and Halle, M. (1968). The Sound Pattern of English. New
York: Harper and Row.
Coker, C.H., Umeda, N., and Browman, C.P. (1973). Automatic synthesis
from ordinary English text. IEEE Audio and Electroacoustics
AU-21, 293-297.
Cooper, W.E. (in preparation). Syntactic Control of Speech Timing.
40
Cooper,W.E. and Ross, J.R. (1975). World order. In R.E. Grossman,
L.J. San, and T.J. Vance (Eds.) Papers from the Parasession
on Functionalism. Chicago: Chicago Linguistic Society. Pp. 63-111.
Delattre, P. (1966). A comparison of syllable length conditioning
among languages. International Review of Applied Linguistics 4,
183-198.
Du Mortier, B. (1963). The physical characteristics of sound emissions
in arthropoda. In R.-G. Busnel (Ed.) Acoustic Behavior of
Animals. Amsterdam: Elsevier. Pp. 346-373.
Fodor, J.A. (1975). The Language of Thought. New York: Thomas Y.
Crowell.
Fodor, J.A., Bever, T.G., and Garrett, M.F. (1974). The Psychology of
Language: An Introduction to Psycholinguistics and Generative
Grammar. New York: McGraw Hill
Fromkin, V.A. (1971). The non-anomalous nature of anomalous utterances.
Language 47, 27-52.
Fry, D.B. (1955). Duration and intensity as physical correlates of
linguistic stress. Journal of the Acoustical Society of America
27, 765-768.
Gay, T., Ushijima, T., Hirose, H., and Cooper, F. S. (1974). Effect of
speaking rate on labial consonant-vowel articulation. Journal
of Phonetics 2, 47-63.
41
Garrett, M.F. (1975). The analysis of sentence production. In G. Bower
(Ed.) Advances in Learning Theory and Motivation: Vol. 9.
New York: Academic Press.
House, A. (1961). On vowel duration in English. Journal of the Acoustical
Society of America 33, 1174-1178.
Huggins, A.W.F. (1969). A facility for studying perception of timing
in natural speech. 2arterly Prgss Report of the M.I.T.
Research Laboratory of Electronics 95, 81-83.
Klatt, D.H. (1975a). Vowel lengthening is syntactically determined
in a connected discourse. Journal of Phonetics 3, 129-140.
Klatt, D.H. (1975b). Structure of a phonological component for a
synthesis-by-rule program. Paper presented at the 90th Meeting
of the Acoustical Society of America, November, 1975.
Klatt, D.H. and Cooper, W.E. (1975). Perception of segment duration in
sentence contexts. In A. Cohen and S.G. Nooteboom (Eds.)
Structure and Process in Speech Perception. Heidelberg:
Springer-Verlag.
Kloker, D. (1975). Vowel and sonorant lengthening as cues to phonological
phrase boundaries. Paper presented at the 89th Meeting of the
Acoustical Society of America, April, 1975.
Kozhevnikov, V.A. and Chistovich, L.A. (1965). Speech: Articulation and
Perception. Joint Publications Research Service 30, 543, U.S.
Department of Commerce, Washington, D.C.
42
Lashley, K.S. (1951). The problem of serial order in behavior. In
L.A. Jeffress (Ed.) Cerebral Mechanisms in Behavior: The Hixon
Symposium. New York: Wiley.
Lehiste, I. (1970). prasegmentals. Cambridge: M.I.T. Press.
Liberman, A.M. (1970). The grammars of speech and language.
Cognitive Psychology 1, 301-323.
Lieberman, P.H. (1967). Intonation, Perception, and Language. Cambridge:
M.I.T. Press.
Lilly, J.C. (1963). Distress call of the bottlenose dolphin: stimuli
and evoked behavioral responses. Science 139, 116-118.
Lilly, J.C. (1965). Vocal mimicry in Tursip: ability to match number
and durations of human vocal bursts. Science 147, 300-301.
Lindblom, B. and Rapp, K. (1973). Some temporal regularities of spoken
Swedish. Papers from the Institute of Linguistics, University of
Stockholm, Publication 21.
Lisker, L. (1974). On "explaining" vowel duration variation. Glossa 8,
233-246.
MacKay, D.G. (1972). The structure of words and syllables: evidence
from errors in speech. Cognitive Psychology 3, 210-227.
MacKay, D.G. (1973). Complexity in output systems: evidence from
behavioral hybrids. American Journal of Psychology 86, 785-806.
MacNeilage, P.F. (1970). Motor control of serial ordering of speech.
Psychological Review 77, 182-196.
43
Malecot, A. (1969). The effect of syllabic rate and loudness on the
force of articulation of American stops and fricatives.
Phonetica 19, 205-216.
Martin, J.G. (1970). On judging pauses in spontaneous speech. Journal
of Verbal Learning and Verbal Behavior 9, 75-78.
Martin, J.G. (1972). Rhythmic (hierarchical) versus serial structure
in speech and other behaviors. Psychological Review 79, 487-509.
Nottebohm, F. (1970). Ontogeny of bird song. Science 167, 950-956.
Ohala, J.J. (1970). Aspects of the control and production of speech.
Unpublished Ph.D. Dissertation, U.C.L.A., Los Angeles, Ca.
Peterson, G.E. and Lehiste, I. (1960). Duration of syllabic nuclei in
English. Journal of the Acoustical Society of America 32, 693-703.
Postal, P.M. (1964). Underlying and superficial linguistic structure.
Harvard Educational Review 34, 246-266.
Reddy, D.R. (Ed.) (1974). Speech Understanding Systems. New York:
Academic Press.
Shattuck, S.R. (1974). Unpublished Ph.D. Dissertation. M.I.T.,
Cambridge, Mass.
Simon, H.A. and Kotovsky, K. (1963). Human acquisition of concepts
for sequential patterns. Psychological Review 70, 534-546.
Thorpe, W.H. (1961). Bird-s Cambridge: Cambridge University Press.
44
von Frisch, K. (1950). Bees, Their Vision, Chemical Senses, and
Language. Ithaca, New York: Cornell University Press.
Winter, P., Ploog, D., and Latt, J. (1966). Vocal repertoire of the
squirrel monkey (Saimiri sciureus). Experimental Brain Research 1,
359-384.
45
Figure Captions
Figure 1: Spectrographic traces of birdsong in the chaf finch
(from Nottebohm, 1970).
Figure 2: A currently popular model of information-processing in
speech production.
0 1 0
0
9
No0
3
I I *1'
VI ~ ~
'I
to N~4~4. ~.
Na
SI
'9,
ml
.9,
S.S
" N I i I
47
Semantic Representation
Deep Structure Representation
Lexical Insertion
Syntactic Transformations
Surface Structure Representation
Word Level Phonological Representation
ISyllable Level Phonological Representation
Phonetic Representation
48
CHAPTER 2
Syntactic Control of Segment Length in Speech
49
Abstract
The durations of segments in spoken sentences were measured
to determine whether speakers alter durations depending on syntactic
structure. The speakers read aloud sentences for which the effects
of three syntactic deletions were assessed in a relatively fixed
phonetic environment. Speakers shortened the durations of segments
just prior to the locations of two deletions which erase material
from the beginning of a clause, whereas speakers lengthened the
durations of segments just prior to the location of a deletion that
erases material from the middle of a clause.
50
The control of timing in speech production is governed by a
variety of factors, including syntactic and phonetic influences. For
example, the duration of a stressed vowel is modified according to its
position in a surface clause (1) and whether it is followed by a voiced
or voiceless consonant (2).
The work to date on syntactic influences of timing has suggested
that major clause boundaries produce lengthening of segments in a pre-
ceding word (1). However, the syntactic effects have not been assessed
independently of phonetic influences on timing, nor has work been directed
at the question of whether one or another type of linguistic representa-
tion corresponds to the level of syntactic processing that mediates
legthening.
A control for phonetic influences was provided here by requiring
speakers to read sentences in which a key word or other segment was
inserted in a relatively fixed phonetic environment. Using this pro-
cedure, it was possible to test two linguistic hypotheses about the
kind of clausal analysis that a speaker computes during sentence
production.
According to one hypothesis, pre-clausal lengthening is determined
at a level of computation comparable to the level of linguistic represen-
tation known as deep, or underlying structure (3). Alternatively,
lengthening could be controlled at a level of processing comparable to
a representation of surface structure. The linguistic rationale for
distinguishing deep and surface syntactic descriptions is well-motivated
(3) and forms a central thesis of generative grammar as well as most other
current grammars that attempt to provide an adequate account of sentence
well-formedness. According to generative grammar, the deep and surface
51
levels of description are mediated by a set of transformations which
permute, add, or delete elements from the underlying structure in
order to derive a surface form. The experiments here were designed to
test whether pre-clausal lengthening is determined by a level of speech
computation corresponding to deep vs. surface linguistic structure.
Experiment I involved sentences like (a) I have the tape (that)
the officer erased, in which a relative pronoun that can be deleted
optionally by a transformation known as Relative Clause Reduction (4).
The deletion of a relative pronoun in Sentence (a) is accompanied by the
deletion of the sentence-node dominating the subordinate clause the
officer erased the tape in underlying structure (5). Thus, if pre-clausal
lengthening is determined by surface rather than underlying structure,
lengthening should be observed for the word tape in Sentence (a) only
when this sentence contains the relative pronoun in surface structure.
The experiment involved two sentences, namely the full (no
deletion) and reduced (deletion) versions of Sentence (a). The key word
tape was bound on the left by the same word the in both sentences and
on the right by the words the and that, containing the same initial
phonetic segment [&Iand stress (6).
Sixteen speakers read each of these two sentences 6 times in
succession, beginning with either the full or reduced versions, according
to random assignment. The speakers were seated in a sound-insulated
chamber and were told at the beginning of the test to practice the
sentences so as to be able to read the sentences "as unitary wholes"
rather than "word-by-word" as in unpracticed reading. In addition,
the speakers were instructed not to place contrastive or emphatic
stress on any words in the sentences.
All utterances were recorded onto magnetic tape via a Neumann
U87 microphone and a Revox A77 tape recorder. The duration of the key
52
word tape was measured from digitized oscillographic traces (sampling
rate = 10 kHz) of the first five occurrences of each sentence for each
speaker. The segment duration for tape excluded the It] and [p]
closure intervals and a [p] release (if any existed) in order to
minimize variability. The segment boundaries were marked with the
aid of a computer controlled cursor (7), and the accuracy of each
measurement was estimated to be within + 2 msec.
The results for the 16 speakers showed a small but statistically
significant difference in the duration of the key word tape (p<.02,
t = 2.755, df = 15; two-tailed t-test for matched pairs). The average
duration of the sentence containing the relative pronoun was 6.1 msec
longer than the duration of the reduced sentence. The results support
the hypothesis that pre-clausal lengthening is controlled in part by a
level of computation that corresponds to a surface rather than deep
level of linguistic description. The results are paralleled by a recent
study of speech perception which showed a stronger clause-boundary
influence on click location in the case of full vs. reduced relative
clauses (8).
Experiment II was designed to test the generality of the above
effect with a second deletion that also operates to delete material from
the beginning of a clause as well as deleting a sentence-node. This
deletion, Conjunction Reduction, deletes material optionally under
identity with material in another clause, as in Dave rehearsed the show
on Tuesday and (Dave) taped it the same day. If the results of Experiment
I depended on surface structure relations, rather than on the particular
form of the deletion rule (identity vs. non-identity deletion), then the
53
effect obtained in Experiment I should also be observed with
Conjunction Reduction.
Experiment II involved 18 speakers, 16 of whom served in
Experiment I. Three sentences were used: (a) Dave rehearsed the -
show on Tuesday and Ted taped it the same day; (b) Dave rehearsed
the sh-w on Tuesday and taped it the same day; and (c) Dave rehearsed
the show on Tuesday and taped it the very same day. Sentence (a)
represented a full coordinate structure, whereas Sentences (b) and
(c) were derived by Conjunction Reduction. Sentence (c) was included
to test the possibility that the lengthening effect for full vs. deleted
versions is conditioned by the total length of the sentences. The
segment chosen for measurement in each of the three sentences included
the final syllable of Tuesday, the word and, and any pause existing in
between.
The results showed that the durations of this segment were
significantly longer for the full coordinate structure of Sentence (a)
than for either of the reduced Sentences (b) and (c), as predicted on
the basis of the surface structure hypothesis [(a) vs. (b): p<.05,
t = 1.982, df = 17; (a) vs. (c): p<.05, t = 1.926, df = 17; one-tailed
t-tests for matched pairs]. The average duration of the key segment in
Sentence (a) was 27.6 msec longer than in Sentence (b) and 39.8 msec
longer than in Sentence (c). No significant difference was observed
between the two reduced Sentences (b) and (c) (p>.20, t = 0.938, df = 17).
The effects obtained in this experiment add further support to the notion
that pre-clausal lengthening is determined at least in part by clausal
relations that exist at a surface rather than underlying level of des-
cription. The results of this experiment, however, are also compatible
54
with an account of timing in terms of rhythmic stress patern (9).
Experiment III was conducted to test the possibility that a
deletion rule which erases material from the middle of a clause and is
not accompanied by sentence-node deletion would also produce lengthening
at the site of deletion. If so, then the previous effects would need
to be regarded as being determined by the process of syntatic deletion
per se rather than by the particular surface structure relations that
hold as a consequence of deletions that erase material from the
beginning (or presumably, the end (10) of a clause and delete an under-
lying sentence-node.
Gapping (11) is a rule that optionally deletes a verb from the
middle of a clause under identity with the verb of another clause,
leaving the general structure of both clauses intact. Two sentences
were used in Experiment III: (a) The lecture began before lunch and the
test began before three; and (b) The lecture began before lunch and the
test before three o'clock. Sentence (a) involved no deletion, whereas
Sentence (b) was derived via Gapping. Both sentences contained the
same number of words and syllables. Measurements of duration were made
on the key word test.
The results showed that the average duration of the key word in
the reduced Sentence (b) was 8.6 msec longer than the duration of this
same word in Sentence (a). This trend was opposite in direction from
the lengthening effects observed for the non-deletion versions of
sentences in Experiments I and II. The effect observed in Experiment
III did not reach statistical significance (.20> p>.10, t = -1.492,
df = 15; two-tailed t-test for matched pairs). The results of this
experiment show that the effects observed in the previous experiments
55
must be accounted for by a level of processing that is sensitive to
the clause structure relations that exist in surface structure. Other
experiments using the sentence-reading procedure also support this
claim (12).
The magnitude and consistency of the lengthening effects for
some speakers was sufficiently great to warrant the implementation of
such effect in programs to synthesize speech by machine from a phonetic
transcription (13).
56
References and Notes
1. J.G. Martin, J. Verb. Lrn. Verb. Beh. 9, 75 (1970); B. Lindblom
and K. Rapp, Papers from Inst. Ling., Univ. Stockholm, Publ. 21
(1973); D.H. Klatt, J. Phonetics 3, 129 (1975).
2. G.E. Peterson and I. Lehiste, J. Acoust. Soc. Amer. 32, 693
(1960); A. House, J. Acoust. Soc. Amer. 33, 1174 (1961).
3. N. Chomsky, Syntactic Structures (Mouton, The Hague, 1957);
P.M. Postal, Harvard Ed. Rev. 34, 246 (1964); N. Chomsky, Aspects
of the Theory of Syntax (MIT Press, Cambridge, 1965); N. Chomsky,
in A Festschrift for Morris Halle, S.R. Anderson and P. Kiparsky,
Eds. (Holt, Rinehart, and Winston, New York, 1973), pp. 232-285.
4. C.S. Smith, Lang. 40, 37 (1964).
5. J.R. Ross, in Modern Studies in English, D.A. Reibel and S.A. Schane,
Eds. (Prentice-Hall, Englewood Cliffs, N.J., 1969), pp. 288-299.
Ross's original formulation of sentence-node pruning was motivated
for the form of relative reduction known as whiz deletion (4),
in which the relative pronoun and verb are deleted. However, the
argument involving extraposition which Ross used as motivation
for sentence-node pruning in the case of whiz deletion also applies
to cases in which the relative pronoun alone is deleted (J. Hankamer,
Constraints on Deletion in Syntax, Ph.D. Dissertation, Yale Univ.,
1971).
6. To the extent that slightly more stress was placed on that vs. the
in the two sentences, a phonetic influence of lengthening on the
key word would be predicted ia the direction opposite that
predicted on the basis of the syntactic hypothesis (A.W.F. Huggins,
Quart. Prog. Rpt. M.I.T. Res. Lab. Electr. 114, 179 (1974).
57
7. A.W.F. Huggins, Quart. Prog. Rpt. M.I.T. Res. Lab. Electr. 95,
81 (1969).
8. J.A. Fodor, J.D. Fodor, M.F. Garrett, and J.R. Lackner, Quart.
Prog. Rpt. M.I.T. Res. Lab. Electr. (1976), in press. The
reduced relatives in their work differed from those of the present
study in that they were derived by whiz deletion (see Note 5).
9. Note that the key segment of the undeleted sentence was followed
immediately by two strongly stressed words, whereas the correspond-
ing segment of the deleted sentences was followed by a single
stressed word. It is conceivable that speakers lengthened the
duration of the key segment in the undeleted sentence because
this segment terminated a rhythmic phrase group. Such an explana-
tion, however, cannot account for the difference obtained in
Experiment I and wrongly predicts other results obtained with
the sentence-reading paradigm (12).
10. Conjunction Reduction can operate to delete material at either the
beginning or end of a clause.
11. J.R. Ross, in Progress in Linguistics, M. Bierwisch and K.E.
Heidolph, Eds. (Mouton, The Hague, 1970); R.S. Jackendoff,
Ling. Ing. 2, 21 (1971).
12. W.E. Cooper, Syntactic Control of Timing in Speech Production,
Ph.D. Dissertation, M.I.T., 1976).
13. C.H. Coker, N. Umeda, and C.P. Browman, IEEE Aud. and Electroacoust.
AU-21, 293 (1973).
14. Supported by a National Science Foundation Graduate Fellowship
and NIH Grant HD-05168.
58
CHAPTER 3
Syntactic Control of Timing in Speech Production:
a Study of Complement Clauses
59
Abstract
A sentence-reading procedure was used to study the influence
of syntactic structure on the timing of syllables in speech production.
In each sentence, easily segmentable key words were placed in the
environment of a putative syntactic boundary. The immediate phonetic
environments of the key words were held fixed or independently assessed
for their effect on syllable timing. In the main experiment, the
effects of complement clause boundaries were tested in sentences
containing the verbs expected and persuaded. These verbs were used
because they trigger structurally distinct complement types. Each of
a group of 15 speakers read aloud a set of 5 sentences 6 times each.
The waveforms of the first 5 occurrences of each sentence were digitized
at a sampling rate of 10 kHz and measured for the duration of the last
two syllables of the verb and for the duration of the immediately
following noun. The results showed that the last two syllables of the
verb expected but not of persuaded were significantly shorter, by a
small amount averaging about 12 msec, when the verb was followed by a
complement clause appearing in the surface structure of the sentence as
compared with when the verb was followed by a non-surface complement.
The results support the notion that a level of syntactic computation
controls timing relations in speech production and that this level com-
putes a clausal representation similar to a variant of linguistic
surface structure. Other experiments showed that (1) a phonetic
effect on syllable timing, conditioned by the presence of a following
voiced vs. voiceless segment, extends across a verb-noun phrase boundary
and (2) the difference noted for the verbs expected and persuaded
generalizes at least in part to other verbs in sentences having the
same structural distinction.
60
1. Introduction
The control of timing in speech production is governed by a
variety of factors, including overall speaking rate as well as phonetic
and syntactic influences. For example, the duration of a stressed
vowel differs depending on whether a following consonant is voiced vs.
voiceless (Peterson and Lehiste, 1960; House, 1961; Delattre, 1966),
whether the vowel is contained in an emphasized word (Lieberman, 1967;
Bolinger, 1972), or whether the vowel is contained in a clause-final
vs. non-clause-final syllable (Martin, 1970; Lindblom and Rapp, 1973;
Huggins, 1974; Klatt, 1975; Kloker, 1975).
The syntactic influences on speech timing may provide an
opportunity to examine the kinds of computations that a speaker per-
forms during the formation of sentence structure. Currently, very little
is known about such computations because sentence production is not
easily amenable to experimental manipulation. Evidence from speech
errors (see Garrett, 1975 and references cited therein) has been the
most useful source of information heretofore, but an experimental
setting is needed to test hypotheses generated either by speech error
analyses or by linguistic theory.
In this paper, a sentence-reading procedure is used which
allows the investigator to examine some of the effects of syntactic
structure on syllable timing in a controlled yet relatively natural
speaking situation. The sentence-reading procedure has been applied to
a number of syntactic constructions (Cooper, 1975a), and the data ob-
tained with the technique are used to distinguish possible computational
models of sentence production as well as to guide two related research
61
areas--speech synthesis by rule (Coker, Umeda, and Browman, 1973)
and the study of perception of duration and its possible role in
aiding listeners to recover structural information (Klatt and Cooper,
1975).
1.1 Underlying vs. Surface Structure
The research on syntactic influences of timing to date has shown
that syllable lengthening occurs primarily at the ends of major clause
and phrase boundaries,1 where boundary locations are determined by
parsing methods similar to those taught in grade school. However,
modern grammatical work typically distinguishes between two or more
levels of syntactic representation for a given sentence, and studies
of syllable lengthening have not yet been designed to test whether the
lengthening effect is produced by one or another level of syntactic
processing.
Regardless of whether one chooses to adopt a transformational
grammar or one of a number of alternatives, a property common to
virtually all current formulations of grammar is the distinction
between a surface representation of word order and another level or
levels of syntactic, logical, and semantic relations. A variant of
this distinction appears to underlie the computational processes
that mediate speech production, according to a recent analysis of
speech errors by Garrett (1975).
The present study was designed to test whether syllable lengthen-
ing might be sensitive to the distinction between surface and underlying
structure. Complement clauses were chosen for sentence materials, since
62
linguistic hypotheses distinguishing underlying and surface representa-
tions have been particularly explicit in the case of complements
(Rosenbaum, 1967; Chomsky, 1973; Postal, 1974); moreover, as recent
discussions by Chomsky and Postal indicate, the analysis of complement
structures is pivotal in current controversy over the general form
that grammatical rules and constraints on such rules should take (for
a review, see Cooper, 1975b). Thus, in addition to providing a means
of testing whether syllable timing is primarily controlled at an
underlying or surface level of syntactic representation, the study of
complementation affords an opportunity to test the relevance of competing
linguistic analyses for developing a performance model of the kinds
of computations carried out by these processing levels.
Although the clausal analysis of complements does provide a
strong linguistic backdrop to the present study, there is no assurance
a priori that complement clauses are accompanied by syllable lengthening
at all in speech production, unlike other clause types (Cooper, 1975a).
Coordinate clauses, non-restrictive relatives, and conditionals are
marked by a comma in written English and are bounded by perceptible
syllable lengthening and pauses in spontaneous speech; by contrast,
complement clauses are neither accompanied by a comma in writing nor
by a perceptibly obvious terminal lengthening in speech.
1.2 EXPECT vs. PERSUADE: underlying structure
The structure of the complement clauses used in this study
will now be reviewed, beginning with an analysis of underlying structure.
A major distinction between the underlying structures of two types of
63
complement was pointed out by Chomsky (1965) and is illustrated by
sentences (1) and (2), containing the verbs expect and persuade.
(1) The host expected Kate to be at breakfast.
(2) The host persuaded Kate to be at breakfast.
Although these two sentences are quite similar superficially in their
word order, they have quite different logical structures. This
difference is brought out clearly when one considers the meanings of
passive sentences containing expect vs. persuade, as in (3) vs. (4):
(3) The host expected Kate to be brought by an escort.
(4) The host persuaded Kate to be brought by an escort.
Sentence (3) is synonymous with the active sentence The host expected
an escort to bring Kate, whereas sentence (4) is not synonymous with the
corresponding The host persuaded an escort to bring Kate. The logical
distinction underlying this difference is that expect Kate X does not
entail expect Kate, whereas persuade Kate X does entail persuade Kate.
The question of how this underlying distinction between the
complements of expect and persuade is to be represented (i.e., in the
underlying structure of the syntactic component of grammar or in a
separate component marking logical relations) remains to some extent
problematic. If the distinction is represented in the syntactic
component, according to the proposal of Rosenbaum (1967), then the
complments of the two verbs are assigned different clause status. Thus,
to the extent that the underlying clause representation of complements
determines syllable lengthening, the complements of expect vs.
persuade should produce an observable difference in speech timing.
In particular, Rosenbaum proposed that sentences like (1)
and (2) have underlying structures like those shown in Figure 1.
64
Insert Figure 1 about here
Under this proposal, the critical distinction between the complements
is that the expect complement is immediately dominated by a Noun Phrase
node in underlying structure, whereas the persuade complement is im-
mediately dominated by a Verb Phrase node. As a consequence, Kate is
a member of the superordinate clause in the structure of the persuade
sentence but is a member of the subordinate clause in the structure
of the expect sentence. Put another way, the major clause boundary for
the persuade sentence occurs immediately after Kate, whereas the major
boundary for the exect sentence occurs immediately before Kate, just
after the verb expect. If the underlying clause structure of complements
determines syllable lengthening, then lengthening should be observed
for Kate in sentence (2) and for expect in sentence (1), relative to
some reference duration.
1.3 EXPECT vs. PERSUADE: surface structure
The underlying structures shown in Figure 1 can be converted
into surface structures by application of transformational rules--rules
which move, add, or delete elements (Chomsky, 1965). For a verb
like expect, the underlying structure can be converted into two
types of surface complement--an infinitival complement (introduced by
to), as in sentence (1), or a that complement (introduced optionally by
that), as in sentence (5) below.
(5) The host expected (that) Kate would be at breakfast.
Generative grammarians agree that the surface structure of that complements
65
like (5) preserves the constituent structure of the underlying representa-
tion shown in Figure 1. In bracketed form, the constituent structure
of (5) can thus be represented by (5'):
(53) [I[[NPthe host][VPexpect]][ [NPKate] [Vbe at breakfast]]]
According to this analysis, there is a major clause boundary immediately
after the verb epected in sentence (5) both in surface and underlying
structure.
For infinitival complements with expected, however, there exist
two major alternative ways of describing the surface representation.
According to one alternative, advocated principally by Rosenbaum (1967)
and Postal (1974), the surface structure of infinitival complements
differs from its underlying structure, and hence, differs from the
surface structure of that complements as well. Rosenbaum and Postal
propose that the noun Kate in infinitival complements like (1) has
been moved ("raised") from its position as the subject of the subordinate
clause in underlying structure (see Figure 1) into the object position
of the higher clause in surface structure. The transformational rule
proposed to account for this movement is termed Raising (from subject
to object). 2
Note that according to the Raising analysis of infinitival
complements, the surface structure of (1) contains a major clause
boundary immediately after the noun Kate; in contrast, the surface
structure of that complements contains a major clause boundary after
the main verb, just prior to Kate. The bracketed surface form of (1)
is represented by (1') below, assuming the Raising analysis:
(1') [S[[NP the host][VPexpect Kate]][ S[NP0 1 EVPbe at breakfast]]]3
The Raising analysis of infinitivals was motivated by the need
66
to account for a number of differences in grammaticality between
infinitival vs. that complements under the application of certain
transformations (e.g., Passive) and constraints on transformations
(e.g., Inclusion Constraint--see Postal, 1974 and Cooper, 1975c for
discussion). However, another plausible alternative account of
infinitival complements has been proposed by Chomsky (1973) to account
for the same range of facts. According to Chomsky's analysis, the
surface structure of (1) is the same as (5) insofar as major constituency
relations are concerned. To account for the variety of differences in
grammaticality between infinitival and that complements, Chomsky
proposes that a distinction be made between finite and infinitival
clauses, and that a universal condition on transformations, the
Tensed-S Condition, be used to account for the differences between
the complement types.
At present, the linguistic controversy surrounding the two
alternative accounts of infinitival complements is not settled.
However, it is of independent interest whether either of these two
proposals provides a better model of performance in sentence production.
This question can be tested in this study of syllabel timing because
the two proposals make conflicting predictions about the location of
syllable lengthening, under the assumption (requiring independent
verification) that a surface as opposed to underlying level of representa-
tion controls such lengthening, at least in part. According to the
Raising analysis, lengthening should appear on the noun Kate of an
infinitival complement with expect but on the main verb of the cor-
responding that complement; in contrast, according to Chomsky's Tensed-S
account, lengthening should appear on the main verb for both types of
complement.
67
We turn now to consider the surface structure of persuade
complements, since the difference between the Raising and Tensed-S
proposals is also represented here. Unlike expect, the verb persuade
takes only infinitival complements with an underlying structure like
that shown in Figure 1. The infinitival sentence in (2), repeated
below for convenience, is the relevant form.
(2) The host persuaded Kate to be at breakfast.
Given the underlying representation of Figure 1, a transformational
rule must be postulated to delete one of the two occurrences of the
noun Kate to convert the underlying structure into a surface form.
Generative grammarians agree for the most part that the correct
formulation involves deleting the subordinate occurrence of Kate,
under identity with the superordinate one, by a rule known as Equi-NP
Deletion. After this rule has applied, the surface structure of (2)
takes the bracketed form shown in (2').4
(2') [ [ [NPthe host][VFpersuaded Kate]][S[NP 0 (VPbe at breakfast]]]
The major clause break of this structure appears immediately after the
noun Kate. Thus, syllable lengthening should be observed for this noun
to the extent that such lengthening is determined by either underlying
or surface structure.
Note that according the the Raising analysis, the major clause
boundary for the persuade infinitival occurs at the same location as
for the expect infinitival in surface structure, namely after Kate. But,
according to Chomsky's analysis, a difference in the boundary locations
for expect vs. persuade infinitivals exists, with the boundary occuring
after the main verb for expect but after Kate for the persuade complement.
By including persuade infinitivals in this study, it was thus possible to
68
provide another test of the merits of the two proposals in accounting
for syllable lengthening in speech.
A summary of the clausal analysis of expect and persuade
complements is provided in Table 1, indicating the major clause
boundaries predicted by the Raising and Tensed-S proposals.
Insert Table I about here
1.4 TO BE deletion
Since the Raising and Tensed-S analyses make conflicting predic-
tions about the locations of clause boundaries for the surface but
not the underlying representations of complements, it is necessary to
try to provide an independent test of whether syllable lengthening is
primarily determined by one or the other of these two levels of
representation. Fortunately, evidence from a variety of other experi-
ments using the same testing procedure as the present study indicates
the presence of syntactic effects which can be attributed to surface
but not underlying structural relations (Cooper, 1975a). However,
none of these other experiments involved complement structures, so a
further test of surface structure effects was desired here.
Consider sentence (6) , which contains no complement clause in
surface structure. This sentence is nearly synonymous with sentence (1)
[The host expected Kate to be at breakfast]. Until recently, it was
(6) The host expected Kate at breakfast.
assumed that (6) and (1) contained identical underlying structures, with (6)
69
being derived by application of a rule which deletes to be. Under
such an analysis, a comparison of sentences like (1) and (6) would
provide the desired independent test of the role of surface vs.
underlying structure as a determinant of syllable lengthening. Borkin
(1973) has shown that a slight difference in meaning is usually
associated with sentences (1) vs. (6), however, such that the version
in which to be has been deleted has a greater tendency to denote
personal experience on the part of the subject. Sentences (7) and (8)
appear to bring out Borkin's point clearly.
(7) I find this chair to be uncomfortable. = Borkin's (lOb)
(8) I find this chair uncomfortable. = Borkin's (10c)
Borkin notes that either (7) or (8) would be appropriate for a speaker
who is reporting on his personal experience with the chair, whereas
only (7) would be approrpiate for a speaker who is reporting the results
of a consumer reaction test in which he himself did not have experience
with the chair. Since both (7) and (8) can be used in the former
circumstance, however, it is still maintainable that sentence pairs
like (l)-(6) and (7)-(8) are derived from the same underlying structure
5
on their most common reading. Assuming this analysis, if syllable
lengthening is controlled primarily by surface as opposed to underlying
structure in the case of complements, lengthening should be observed
for the verb expected in (1) vs. (6), since only in the former
sentence does the underlying complement clause exist in surface
structure. For (6), the subordinate sentence node in underlying
structure is presumably deleted by the condition of S-node pruning
(Ross, 1969; Hankamer, 1971; Reis, 1973), a convention which, according to
its original formulation, deletes all S-nodes from surface structure
70
which do not branch into a verb phrase and some other constituent.
Unlike expect, the optional deletion of to be is not permitted
for the complement of a verb like persuade, as evidenced by the major
difference in meaning between a sentence like (2) and (7) below.
(7) The host persuaded Kate at breakfast.
Unlike (2) [The host persuaded Kate to be at breakfast], (7) is
assumed to contain a single clause in underlying as well as surface
structure. In addition to the surface structure contrast between
expect sentences like (1) and (6), it was decided to include the con-
trast between persuade sentences like (2) and (7) in the present study.
Because (2) and (7) differ greatly in their meaning as well as in their
constituent structure at both underlying and surface levels, however,
it was not possible to make any firm predictions about the effects of
this comparison on syllable timing. The absence of prediction in
this case resulted from a lack of prior knowledge about any effects
that semantic representation might have on timing.
2. A Sentence-Reading Paradigm
In order to study a speaker's computational processes during
speech production, it would be most desirable to study a corpus of
spontaneous speech. While certain effects, including very general
effects of syntactic structure on syllable timing, can indeed by
studied in this manner (Kloker, 1975), work on specific hypotheses
cannot yet be conducted with proper control or efficiency using such
a data base. In the case of complement clauses, the need for a controlled
experimental setting is particularly acute, since the unaided ear
71
suggests that, to the extent that syllable lengthening effects exist
at all for complements, these effects must be quite small compared
with the lengthening effects exhibited for some other clause types,
such as coordinates and non-restrictive relatives. The small magnitude
of the complement effects would presumably be impossible to ascertain
from a corpus of spontaneous speech, since the interplay of other
influences on syllable timing (particularly speaking rate and phonetic
environment) would mask any systematic effects of complement structure.
Accordingly, a sentence-reading procedure was used in this and
similar studies, in whcih the phonetic form of the utterances was
tightly controlled. In each test, a short list of sentences was read
by a speaker after a period of practice during which he was familiarized
with the task and the particular sentence materials. Each sentence
contained one or more key words, placed at the location(s) of putative
syntactic boundaries. The choice of key words was influenced by the
degree to which their phonetic representations could be readily segmented
from a digitized oscillographic trace of the speech waveform. In most
of the studies conducted with the procedure, the key words were CVC
monosyllables containing a long vowel (Cooper, 1975a),6 bounded by
stop consonants or fricatives (e.g., Kate, tape, case). In some
instances, including the present study, key words were required which
were slightly more difficult to segment (e.g., expected, persuaded),
although in each case the additional measurement error incurred with
such words was small compared with the magnitude of the effects observed.
Wherever possible, the key words were placed in the middle of
the sentence string in order to neutralize any possible influence of
differences in subglottal pressure on segmental timing (Lindblom and
72
Rapp, 1973). The sentences themselves typically contained the same
number of words and syllables and shared as many words as possible
with one another, compatible with signalling the structural differences
under study.
Each speaker was tested individually in a sound-insulated
chamber forapproximately 40 minutes per session. During this time,
the speaker was typically tested on from 5 to 7 sentence lists, each
representing a separate experiment. Each of the experiments with com-
plements reported here was conducted as part of a different session.
At the beginning of a session, the speaker was told that the general
purpose of the experiments was to study the syntactic control of
syllable timing. The speaker was then told that he would be given
practice in reading lists of sentences in order to train him to
utter each sentence "as a unitary whole, as if it were spoken spontane-
ously", rather than "word-by-word", as in unpracticed reading.7 The
speaker was encouraged to speak as naturally as possible but to avoid
placing contrastive or emphatic stress on any word or syllable in
a sentence.
Following these preliminary instructions, the speaker was
given the first of a series of lists of sentences to practice until
he figured out how he intended to utter each sentence in the list.
The speaker was told to consider each sentence in the list independently
of the others. Following practice, the speaker was asked to utter
each sentence aloud once, providing the experimenter with a final
opportunity to check for undesirable contrastive stress or emphasis
and to check recording levels. The speaker then uttered each sentence
in the list 6 times in succession for recording. The speaker was
73
asked to say each token of the sentence at the same overall rate and
with the same rhythm.
The speaker was told to expect to produce a few mispronunciations
during the test. For mispronunciations, or for any changes from the normal
intonation or timing which the speaker chose to adopt, the speaker was
instructed to pause, utter the word "repeat", and then say the sentence
token again, as often as needed until 6 appropriate occurrences of the
sentence were produced. After reading a given list, the speaker was
provided a short rest period, encouraged to take a drink of water, and
then asked to begin the practice procedure with a new list.
3. Experiment I
In this experiment, the durations of syllables in 5 sentences
were measured to determine whether differences in the syntactic
structure of complements following the verbs expected and persuaded
would produce differences in syllable timing. The verb and the following
noun of each sentence were chosen as the key words for measurement,
since the locations of the major syntactic boundaries postulated
according to both major linguistic analyses of complementation occurred
immediately after one of these two words. The words immediately surround-
ing the verb in each sentence were held fixed, so that any durational
effects observed for the verb could be attributed directly to syntactic
structure. For the noun, the following word, of necessity, covaried
with the syntactic structure of the sentence. For this reason, an
independent test (Experiment II) was required to determine whether
any effects of syllable duration observed for the noun in this
74
experiment were due to phonetic as opposed to syntactic influences.
3.1 Method
Subjects
Fifteen M.I.T. students and employees served as volunteers in
the experiment. The subjects were native speakers of English with
no history of speech or hearing impairment.
Sentence Materials
The sentences used in the experiment are listed below, along
with their descriptive labels (written in capital letters):
(a) The host expected Kate to be at breakfast. (EXPECT-INF = expect
infinitival complement)
(b) The host expected Kate would be at breakfast. (EXPECT-THAT =
expect that complement)
(c) The host expected Kate at the big breakfast. (EXPECT-SIMPLE =
expect single surface clause)
(d) The host persuaded Kate to be at breakfast. (PERSUADE-INF =
persuade infinitival complement)
(e) The host persuaded Kate at the big breakfast. (PERSUADE-SIMPLE =
persuade single surface clause)
Each of these 5 sentences contained 8 words and 11 syllables. When read
without contrastive or emphatic stress, the sentences also had the same
approximate sentence stress countour, with primary stress on Kate.8
Sentences (a) and (d) were equivalent to sentences (1) and (2)
discussed in the Introduction. Sentences (c) and (e) had the same structure
75
as (6) and (7), while sentence (b) represented a version of (5) in
which the that complementizer had been deleted. The complementizer
was deleted in (b) so that the immediate phonetic environment of the
verb would be identical for all 5 sentences.
Procedure
The testing procedure was described in Section 2. Each speaker
read each sentence in the list (a)-(e) 6 times, beginning with sentence
(a) or (e), according to a randomized assignment. The first 5 occur-
rences of each sentence, excluding false starts, mispronunciations, and
sentences containing contrastive or emphatic stress, were digitized at
a sampling rate of 10 kHz on a PDP-9 computer at the M.I.T. Research
Laboratory of Electronics. The waveforms were analyzed for segment
durations with the aid of a computer controlled cursor (Huggins, 1969).
This cursor was maneuvered by velocity and position dials to mark the
onset and offset of the bisyllabic segments [spektad] and [swedad] of
the verbs expected and persuaded as well as of the monosyllable [ket]
(Kate). The time difference between the segment onset and offset
was displayed on an oscilloscope screen to the nearest 100 isec.
The onset of visible frication following the closure gap of
the first [k] in expected was taken as the onset of the final two
syllables of this word.9 For some speakers, a [k] release burst was
discernible in the waveform and in such cases this burst was not
included as part of the measured bisyllabic segment, since including
the burst duration would have introduced greater variability into
the data.
76
The onset of visible frication following the [r] in persuaded
was similarly taken as the onset of the bisyllabic segment for this
verb. The offset of the bisyllabic segments for both verbs was
measured to be the termination of visible glottal pulsing in the un-
stressed syllable [ad], not including the following closure interval
or any [d] release burst. As above, the decision not to include
portions of the waveform as part of the bisyllabic segment was based
upon the undesirability of increasing the measurement variability.
The reliability of each measurement of the verb segments was
estimated to be within + 3 msec, based on a duplicate set of measure-
ments for a small sample of 20 utterances made without the experimenter's
remembrance of, or reference to, the original measurements taken for
these utterances. All waveform measurements were made by the author.
The onset of the monosyllable Kate was measured at the release
burst of [k], and the offset of the syllable was measured at the
termination of visible glottal pulsing. The onset release burst was
included in the measurements of Kate because this burst was clearly
discernible in the waveforms of all speakers and did not increase the
variability of the data appreciably. The closure interval of the [t]
or any [t] offset burst was not included as part of the duration of
Kate. The reliability of each measurement for Kate was estimated to
be within + 2 msec, based on duplicate measurements of the same
utterances cited above.
77
3.2 Results arid Discussion
Verb Duration
The mean durations of the verb segments for pSected and persuaded
are presented for the individual speakers in Table II. Two-tailed
Insert Table II about here
t-tests for correlated observations were applied to the mean durations
for the set of speakers to determine whether any differences in these
durations were statistically significant. The analysis for the bisyllabic
segment of the verb expected showed that the segment in both Sentence (a)
EXPECT-INF and Sentence (b) EXPECT-THAT were significantly longer
than the segment in Sentence (c) EXPECT-SIMPLE (EXPECT-INF vs. EXPECT-
SIMPLE: t = 2.675, df = 14, p < .01; EXPECT-THAT vs. EXPECT-SIMPLE:
t = 2.487, df = 14, p < .02). The average duration of the bisyllabic
segment for EXPECT-INF was 13.3 msec longer than for EXPECT-SIMPLE,
while the average duration of the segment for EXPECT-THAT was 9.8 msec
longer than for EXPECT-SIMPLE.
The significant lengthening of the verb segment for the sentences
(a) and (b) containing surface complement clauses vs. sentence (c),
which had the same meaning as the other two sentences and presumably
the same underlying syntactic structure as well, suggests that com-
plement clause boundaries in surface structure are among the types of
syntactic boundaries which help to determine segmental lengthening.
As expected, however, the lengthening observed for surface complements
was small compared with the effects observed at the boundaries of
78
other clause types such as coordinates and conditionals (Cooper,
(1975a).
No significant difference was found between the duration of the
verb segment for expected between the two complement types, EXPECT-INF
and EXPECT-THAT (t = 0.856, df = 14, p > .20). The average duration
of the segment in EXPECT-INF was longer than the duration in EXPECT-
THAT by 3.3 msec. According to the Raising analysis of complementation,
whereby a major clause boundary exists immediately after expected
for EXPECT-THAT but not EXPECT-INF, the duration of the verb segment
for EXPECT-THAT should have been longer. Since the data in fact
show a slight trend in the opposite direction, the results provide
no support for a syntactic level of computation in speech production
that corresponds to the surface representations proposed by the Raising
analysis. This finding is paralleled in speech perception by a
recent study of click location in which a test for Raising also failed
to show support for such an analysis (Fodor, Fodor, Garrett, and
Lackner, 1975).
The lack of a significant difference between EXPECT-INF and
EXPECT-THAT is, on the other hand, consistent with the analysis of
complement clause structure proposed by Chomsky (1973), whereby
infinitival and that complements have the same surface as well as
underlying constituent structure. .However, the present data cannot
be taken as very strong support for a performance analog of Chomsky's
analysis, given the possibility of a Type II statistical error.
We now turn to consider the bisyllabic segment durations for
the verb persuaded. These durations were generally about 10-20 msec
shorter than the durations of the corresponding segment of the verb
10
expected.
79
The average duration of the verb segment in PERSUADE-INF was
3.3 msec longer than the average duration of the segment in PERSUADE-
SIMPLE. The difference in duration for the set of subjects was not
statistically significant (t = 1.149, df = 14, p > .20). Acceptance
of the null hypothesis here is consistent with the predictions of
both major analyses of complementation, since no major clause boundary
occurred immediately after persuaded in either the underlying or surface
structure of these two sentences.
We conclude the discussion of the verb duration data by noting
that, according to either of the linguistic analysis of complementation,
it was anticipated that some difference in the duration of the
EXPECT sentences should have occurred, whereas no such differences
should have occurred for the PERSUADE sentences. In fact, the only
statistically significant effect obtained in the experiment was for
the EXPECT sentences, providing support for the notion that syllable
timing is conditioned in part by complement clause structure in a
manner corresponding at a general level to a syntactic clause analysis.
The difference obtained between EXPECT-INF and EXPECT-THAT vs.
EXPECT-SIMPLE provides some evidence that the site of syntactic computa-
tion which controls syllable timing computes a clausal representation
corresponding approximately to linguistic surface as opposed to under-
lying structure. Since a valid test of the Raising account of the
possible difference between EXPECT-INF and EXPECT-THAT rested on
the assumption that syllable lengthening was controlled in part at
a surface as opposed to underlying level of representation, and since
this assumption received some independent support, the absence of
a difference between infinitival and that complements predicted by
80
the Raising analysis in this experiment must be considered problematic
for the view that speakers generally compute a syntactic representa-
tion of complements corresponding to such an analysis.
Noun Duration
The results of the duration measurements for the noun Kate
are presented in Table III. Unlike the durations for the verb
Insert Table III about here
segments, the data for the noun can be compared for a phonetically
fixed environment in the case of sentences (a) vs. (d) only, the
infinitival complements for expected and persuaded. In both sentences,
Kate was preceded by the unstressed syllable [ad] and was followed
by the infinitive to. The results showed that the average duration
of Kate for PERSUADE-INF was 2.5 msec longer than the average duration
of Kate for EXPECT-INF. The difference in duration between these
two test conditions was not statistically significant (t = 1.134,
df = 14, p > .20). The slight trend for the noun following persuaded
to be longer than the noun following expected is consistent with the
notion that speakers compute the structural representations of the two
complements according to Chomsky's (1973) account of infinitivals,
whereby a major clause boundary exists immediately after the noun
for persuaded but after the verb for expected. On the other hand,
the absence of a statsitically significant difference between the two
complement types is also consistent with a Raising analysis, whereby a
81
major clause boundary exists immediately after the verb for both
persuaded and expected. In summary, then, the data for the noun
durations of EXPECT-INF vs. PERSUADE-INF provide no independent basis
for distinyuishing the merits of the Chomsky and Rosenbaum-Postal
proposals as models of speech performance.
In addition to the single comparison between phonetically
stable nouns in EXPECT-INF vs. PERSUADE-INF, other comparisons were
made for the noun duration data. These comparisons revealed some
significant differences, potentially attributable to the effects of
the immediate phonetic environment. In particular, the durations
of Kate before at in EXPECT-SIMPLE and PERSUADE-SIMPLE were signifi-
cantly longer than the durations of Kate before to in EXPECT-INF
and PERSUADE-INF and longer than the duration of Kate before would
in EXPECT-THAT (p < .05 in each case). The average duration of Kate
before at in the SIMPLE sentences was about 15 msec longer than before
to and would in the complements.
4. Experiment II
It is known that the duration of a vowel is longer when it is
followed immediately by a voiced vs. voiceless phonetic segment
occurring within the same word (Peterson and Lehiste, 1960; House,
1961; Delattre, 1966). In addition, Barnwell (1971) has shown that
this effect of phonetic environment is stronger within a word than
across a word boundary, although his data base was not sufficiently
large, nor were his sentence materials sufficiently well-matched, to
test the possibility that the phonetic effect does operate across a
82
word boundary to some extent. The data for the noun Kate in
Experiment I, regarding the difference between simple and complement
sentences, could be accounted for by a similar effect, as opposed
to some unexpected difference in syntactic structure. Of necessity,
the structural differences of interest in Experiment I made it impos-
sible to control for the phonetic environment of the noun Kate, unlike
the verb. In this experiment, an independent test was thus carried
out to examine the possibility of a phonetic effect across word
boundaries of the type that would directly account for the results
for Kate in the previous experiment.
4.1 Method
Subjects
Ten subjects, one of whom (I.B.) served in Experiment I,
participated as volunteers in this experiment. All new subjects
had the same qualifications as the subjects of Experiment I.
Sentence Materials
The following three sentences were used in the experiment.
(a) We skate to the farm.
(b) We skate with the crowd.
(c) We skate at the pond.
Each of these sentences contained 5 monosyllabic words and the same
approximate sentence stress countour in non-emphatic speech. In
addition, the underlying and surface structure representations of
the sentences were identical with regard to major constituent relations,
83
consisting of a subject noun phrase and a verb phrase dominating a
verb and a prepositional noun phrase.
Each sentence contained the key word skate, immediately followed
by a word beginning with [t], [w], or [K], These three segments
corresponded to the three segments immediately following the word
Kate in Experiment I. It was decided to use the above sentences for
the test rather than simple phrase like Kate to, Kate would, and Kate
at (taken directly from the sentences of Experiment I) so as to pre-
serve a sentence context.
Procedure
The procedure for testing described previously was used for
this experiment. Each speaker read each sentence in the list (a)
through (c) 6 times, beginning with sentence (a) or (c), according
to a randomized assignment. The durations of the monosyllabic words
we and skate were measured using the same general procedure as
Experiment I.
4.2 Results and Discussion
The mean durations for the words we and skate are presented
in Table IV. The average durations of we in the three sentences were
Insert Table IV about here
all within 3.5 msec of one another, and the differences among these
durations for the set of subjects were not statistically significant
84
(p > .20 in each case). These results are consistent with the notion
that the duration of a word is not significantly influenced by dif-
ferences in the phonetic structure of another word that is two words
removed from it.
The average duration of skate in the three sentences showed
systematic differences. The word skate was longest preceding at,
somewhat shorter preceding with, and shortest preceding to, covering
an average range of more than 24 msec. Statistical tests showed
that the duration of skate before at was significantly longer than
the duration of skate before to (t = 3.683, df = 9, p < .01) and
that the duration of skate before with was also significantly longer
than the duration of skate before to (t = 7.043, df = 9, p < .001).
On the other hand, the duration of skate before at was not signifi-
cantly longer than the duration of skate before with (t = 1.246,
df = 9, p > .20).
The results of this experiment indicate that the duration
of skate is significantly longer when the following word begins with
either of the voiced segments [a] or [w], in comparison with the
voiceless segment [t]. This pattern of results demonstrates that
the conditioning effect of following voiced vs. voiceless segments
extends across a word boundary, and, furthermore, that it extends
across a relatively major surface phrase boundary separating the
verb and prepositional noun phrase. The magnitude of this effect
is the same order of magnitude as the effects observed in Experiment I.
We can thus conclude that most, if not all, of the segmental lengthening
observed for that noun was attributable to the phonetic structure
of the following segment rather than to the distinction between simple
and complement sentences.11
85
5. Post-Hoc Analysis of Individual Speakers' Data
For Experiment I
Based on the results of Experiment II, a post-hoc analysis
was carried out on the data of Experiment I to determine whether any
individual speakers showed a pattern of results that was strikingly
consistent with either a Raising or Tensed-S analysis of complementa-
tion. If a speaker computed syntactic representations for infinitival
vs. that complements according to a Raising analysis, then the duration
of the verb segment of expected should have been longer for EXPECT-
THAT than for EXPECT-INF. In addition, based on a consideration of
the results of Experiment II, the duration of the noun following
expected in EXPECT-INF should have been longer than the duration of
the noun in EXPECT-THAT, whereas the duration of the noun in EXPECT-
INF should have been about equal to the duration of the noun in
PERSUADE-INF. Two of the 15 speakers of Experiment I, R.F. and M.J.,
showed this pattern of results. Thus, while a Raising analysis
appears incapable of accounting for the results of the speakers as
a group, it is possible that this analysis is represented as a level
of speech computation for two of the 15 speakers. Further research
is required to test the possibility that the particular pattern of
results obtained for these two speakers was not coincidental.
A similar analysis of individual speakers' data was conducted
in search of speakers who showed a pattern of results corresponding
to Chomsky's Tensed-S proposal. As noted earlier, speakers who com-
puted a syntactic representation of infinitival and that complements
according to the Tensed-S account should have produced approximately
86
equal durations for the verb segment of expected in EXPECT-INF
and EXPECT-THAT. Furthermore, taking into account the results of
Experiment II, the speakers should have produced a noun duration for
EXPECT-THAT that was longer than for PERSUADE-INF, while the noun
durations for EXPECT-THAT and PERSUADE-INP should have been about
equal. None of the individual speakers of Experiment I showed this
particular pattern of results. It must thus be concluded that the
post-hoc analysis of individual speakers' data provides no support
for the notion that speakers represent the structure of complements
in a manner like that proposed under a Tensed-S analysis and only
very marginal support for the notion that some speakers represent
the complements in a manner corresponding to a Raising analysis.
6. Experiment III
The single result thus far providing any strong support
for the notion that a surface structure representation of complements
affects syllable timing was the significant lengthening of the verb
segment in Experiment I for the complements EXPECT-INF and EXPECT-
THAT, in comparison with the duration of EXPECT-SIMPLE. To test
whether this difference reflected an improbable chance result or some
idiosyncratic property of the verb expected, it was decided to conduct
a third experiment, similar to Experiment I but using different
verbs.
Unlike the verb expect, there exist some verbs which trigger
the same complement structures as expect but which can also occur as
main verbs in sentences which contain a single underlying clause.
87
Believe is such a verb, and it was used in this experiment in both
single-clause and complement sentences.
6.1 Method
Subjects
Seven M.I.T. students and employees, 5 of whom served in
Experiment I, participated in this experiment. The two new subjects
had the same qualifications as the subjects of previous experiments.
Sentence Materials
The sentences used in this experiment were divided into three
categories, Expect-type 1, Expect-type 2, and Persuade-type 1. The
Expect-type 1 sentences contained comparisons between infinitival
complements and single clause structures from which to be had presumably
been deleted (cf. Borkin, 1973). Four different verbs were used,
each triggering complements in the same manner as expect with regard
to the major constituency relations of relevance (Bresnan, 1972).
The resulting list of 8 sentences appears below:
Expect-type 1
(a) We believed Kate to be crazy.
(b) We believed Kate crazy at times.
(c) We considered Kate to be crazy.
(d) We considered Kate crazy at times.
(e) The boss wants Ted to be at the station by 3 o'clock.
(f) The boss wants Ted at the old train station by 3 o'clock.
88
(g) The boss needs Ted to be at the station by 3 o'clock.
(h) The boss needs Ted at the old train station by 3 o'clock.
The list of Expect-type 2 sentences included three-way comparisons
among infinitival complements, that complements, and single clause
structures considered to contain a single clause in underlying as
well as surface structure. Three different verbs triggering com-
plements like expect were used, making a total of 9 sentences listed
below.
Expect-type 2
(i) Kate believed John to be at the trial.
(j) Kate believed John was at the trial.
(k) Kate believed John at the trial last week.
(1) Kate understood John to be at the trial.
(m) Kate understood John was at the last trial.
(n) Kate understood John at the trial last week.
(o) John proved Bayes' Theorem to be applicable to my
sampling problem.
(p) John proved Bayes' Theorem was applicable to my samplinc
procedure.
(q) John proved Bayes' Theorem at the conference on statisti
procedures.
Finally, the list of Persuade-type 1 sentences included two-way com-
parisons between infinitival complements and single-clause sentences
considered to contain a single clause in both underlying and surface
structure. These sentences used two verbs which trigger complements
like persuade (Bresnan, 1972) .
3
ical
89
Persuade-type 1
(r) We convinced Kate to be at breakfast.
(s) We convinced Kate at the big breakfast.
(t) Ted challenged Kate to be at breakfast.
(u) Ted challenged Kate at the big breakfast.
In summary, the Expect-type 1 and Persuade-type 1 sentences were
comparable to sentences used for expect and persuade in Experiment I.
The Expect-type II sentences, however, contained verbs of the expect
type in single clause sentences of the same kind as the sentences
with persuade type verbs.
Procedure
The same testing procedure was used as in previous experiments.
The sentences appeared on three separate sentence lists. The dura-
tions of the verbs were measured using the general technique described
earlier. Unlike Experiment I, the duration of the entire verb
segment, excluding terminal gaps and offset bursts, were measured.
6.2 Results and Discussion
Expect-type 1
For the 4 Expect-type 1 verbs, 3 verbs showed an average
lengthening effect in the infinitival complement vs. single-clause
sentences for the set of 7 speakers. The average lengthening effect
for believe was 20.1 msec, for want 5.1 msec, and for need only 0.2 msec.
For the verb consider, an average shortening of 7.9 msec was obtained,
although, unlike the date for the other verbs, the average effect for
90
consider was heavily influenced by the effect for a single speaker.
We conclude from these results that the effect observed in Experiment I
generalizes to some extent to other verbs in the context of synonymous
complement vs. single-clause sentences, although a much larger data
base would be required to test generality in a definitive manner.12
Expect-type 2
In contrast to the trend observed for the data of Expect-type 1
sentences, the duration of the main verb for Expect-type 2 sentences
was either shorter for the complement than single-clause structures
or was about equal. For the verb believe, which showed a sizable
lengthening effect for the complement in Expect-type 1 sentences,
an equally sizable lengthening effect was observed for the single-
clause sentence when this verb occurred in an Expect-type 2 context.
The average duration of the verb in the single-clause sentence for
believe in the latter context was 27.0 msec longer than the duration
of the verb with an infinitival complement and 27.4 msec longer than
the duration of the verb with a that complement. An average lengthening
effect for the single-clause sentence was also observed for understood,
amounting to slightly more than 12 msec in comparison to each com-
plement. For the verb proved, however, the verb of the single-clause
was shorter than the verb in the infinitival complement by 11.3 msec
and was longer than the verb in the that complement by 3.4 msec.
In summary, it appears that the verb duration data for Expect-type 2
sentences differs from that of Expect-type 1 sentences in the predicted
direction. The lengthening for the single-clause sentences of
Expect-type 2 is unaccountable in terms of clausal analysis but may
91
well reflect semantic differences of focus that will need to be
studied further. Since the present data base is small and subject
to idiosyncrasies, we will not pursue a discussion of the possible
significance of the single-clause lengthening of Expect-type 2
sentences here. 1 3
Persuade-type 1
The data for the two Persuade-type 1 verbs, convinced and
challenged, showed approximately equal average durations for the
single-clause and complement sentences. The average difference in
duration was within 2 msec for both verbs. These results indicate
that the results for persuaded in Experiment I generalize to these
other verbs of the same structural classification.
7. General Discussion
The present experiments have provided evidence for the
existence of both syntactic and phonetic effects on syllable timing.
At a syntactic level, speakers lengthened the duration of the last
two syllables of the verb expected when this verb was followed by a
complement clause in the surface structure of the sentence, as com-
pared with when it was followed by a simple phrase in surface structure
which, according to one linguistic analysis, was derived from a full
complement in underlying structure. This finding suggested that a
speaker's surface structure representation of an utterance is the
primary level of syntactic representation that exercises control
over syllable timing.
92
The effect observed for full vs. reduced complements is
noteworthy also because the verb in both sentences occurred two words
prior to the point in the sentences at which the distinction between
the surface structures became apparent. This fact suggests that
the speakers in this study were computing the durations of syllables
in part as a function of the hierarchical constituent structure of
the sentences rather than on a linear word-by-word basis.
The major negative finding of the present study was the failure
to find evidence in strong support of a system of syntactic computation
that corresponded to either the Raising or Tensed-S analyses of com-
plementation. As in the case of previous work in psycholinguistics,
it is perhaps not surprising that this study failed to show a close
correspondence between the kinds of structural analyses employed by
a speaker during sentence production and the kinds of analysis proposed
on independent linguistic grounds (see discussion in Fodor and Garrett,
1966 and in Fodor, Bever, and Garrett, 1974). Although it remains
a reasonable research strategy to test the relevance of linguistic
theory to a performance model of sentence production, greater reliance
must be placed in constructing a model of performance based primarily
upon available psycholinguistic data and other psychological considera-
tions. The main problem facing this approach is that the data re-
quired are for the most part unavailable in order to verify even the
most basic of assumptions needed as a foundation for further hypothesis
testing.
Aside from the long-range goal of providing a model of the compu-
tational processes involved in speech production, the results of
studies like the present serve an immediate function as a guide for
93
research on speech syntehsis by rule and on the perception of segmental
timing. The present results indicate that a speech synthesis by
rule program can comfortably avoid any specialized rules for com-
plement clause boundaries, since the average effects are comparable
to the perceptual tolerance of listeners (Huggins, 1972; Klatt and
Cooper, 1975). At most, a 4-5% lengthening of syllables in the
verb preceding a complement clause boundary might be implemented.
Since it appears unlikely that the lengthening found in speech produc-
tion could be detected reliably by listeners in normal speech situa-
tions, there is no need to conduct a systematic study of the perceptual
relevance of segmental lengthening with complements, unlike the case
for some other types of lengthening which are syntactically determined
(Cooper, 1975a).
It is tempting, finally, to speculate on the precise cause
of the syllable lengthening produced by the presence of a surface
clause boundary. Perhaps, if speakers compute articulatory commands
on a clause-by-clause basis to some degree, lengthening is produced
to provide the speaker with an extra fraction of time during which to
plan the articulatory commands for the upcoming clause. On the other
hand, it is conceivable that speakers lengthen syllables to provide
a slightly longer interval over which to place a rise or fall in
intonation associated with the clausal structure. This possibility
is not considered very likely, but a systematic study of fundamental
frequency contours remains to be conducted.
94
Acknowledgements
This article is a version of a chapter of my Ph.D. disserta-
tion (Cooper, 1975a). The work was supported by a National Science
Foundation Graduate Fellowship and NIH Grant HD-05168. I am
irretrievably indebted to Merrill Garrett for guidance throughout
the course of this work. In addition, I thank Dumont Billings, Ann
Cutler, Jerry Fodor, A. W. F. Huggins, Dennis Klatt, Steven Lapointe,
Cornelia Parkes, and John Robert Ross for advice and assistance.
95
References
Akmajian, A. and Jackendoff, R.S. (1970). Coreferentiality and
stress. Linguistic Inquiry 1, 124-126.
Barnwell, T.P. (1971). An algorithm for segment durations in a reading
machine context. Technical Report No. 479. Research Laboratory
of Electronics, M.I.T., Cambridge, Ma.
Bolinger, D. (1972). Accent is predictable (if you' re a mind-reader).
Language 48, 633-44.
Borkin, A. (1973). To be and not to be. In Papers from the Ninth
Regional Meeting of the Chicago Linguistic Society (C. Corum,
T.C. Smith-Stark, and A. Weiser, Eds.). Chicago: Chicago Linguistic
Society. Pp. 44-56.
Bresnan, J. (1971). Sentence stress and syntactic transformations.
Language 47, 257-81.
Bresnan, J. (1972). The Theory of Complementation in English Syntax.
Ph.D. Dissertation, M.I.T., Cambridge, Ma.
Chomsky, N. (1965). Aspects of the Theory of Syntax. Cambridge:
MIT Press.
Chomsky, N. (1973). Conditions on transformations. In A Festschrift
for Morris Halle (S.R. Anderson and P. Kiparsky, Eds.). New York:
Holt, Rinehart, and Winston. Pp. 232-285.
Clark, H.H. (1973). The language-as-fixed-effect fallacy: a critique
of language statistics in psychological research. Journal of
Verbal Learning and Verbal Behavior 12, 335-359.
Coker, C.H., Umeda, N., and Browman, C.P. (1973). Automatic synthesis
from ordinary English text. IEEE Audio and Electroacoustics
AU-21, 293-7.ft
96
Cooper, W.E. (1975a). Syntactic Control of Timing in Speech Production.
Ph.D. Dissertation, M.I.T., Cambridge, Ma.
Cooper, W.E. (1975b). Why reordering transformations? a review of
Paul M. Postal's On Raising. Forthcoming.
Cooper, W.E. (1975c). Inclusion rleations in language and perception.
Forthcoming.
Delattre, P. (1966). A comparison of syllable length conditioning
among languages. International Review of Applied Linguistics
4, 183-98.
Fodor, J.A., Bever, T.G., and Garrett, M.F. (1974). The Psychology
of Language: An Introduction to Psycholinguistics and Generative
Grammar. New York: McGraw-Hill.
Fodor, J.A., Fodor, J.D., Garrett, M.F., and Lackner, J.R. (1975).
Effects of surface and underlying clausal structure on click loca-
tion. Quarterly Progress Report of the M.I.T. Research Laboratory
of Electronics, in press.
Fodor, J.A. and Garrett, M.F. (1966). Some reflections on competence
and performance. In Pscholin tic Pars (J. Lyons and R.J.
Wales, Eds.). Edinburgh: Edinburgh University Press. Pp. 135-179.
Garrett, M.F. (1975). The analysis of sentence production. In
Advances in Learning Theory and Motivation, Vol. 9 (G. Bower, Ed.).
New York:' Academic Press, in press.
House, A. (1961). On vowel duration in English. Journal of the
Acoustical Society of America 33, 1174-1178.
Huggins, A.W.F. (1969). A facility for studying perception of timing
in natural speech. Quarterly Progress Report of the M.I.T.
Research Laboratory of Electronics 95, 81-83.
97
Huggins, A.W.F. (1972). Just-noticable difference for segment duration
in natural speech. Journal of the Acoustical Society of America
51, 1270-1278.
Klatt, D.H. (1975). Vowel lengthening is syntactically determined
in a connected discourse. Journal of Phonetics 3, 129-140.
Klatt, D.H. and Cooper, W.E. (1975) Perception of segment duration in
sentence contexts. In Structure and Process in Speech Perception
(A. Cohen and S.G. Nooteboom, Eds.). Heidelberg: Springer-Verlag.
Huggins, A.W.F. (1974). An effect of syntax on syllable timing.
Quarterly Progress Report of the M.I.T. Research Laboratory of
Electronics 114, 179-185.
Kloker, D. (1975). Vowel and sonorant lengthening as cues to phono-
logical phrase boundaries. Paper presented at the 89th Meeting
of the Acoustical Society of America, April, 1975.
Lieberman, P. (1967). Intonation, Perception, and Language. Cambridge:
MIT Press.
Lindlbom, B. and Rapp, K. (1973). Some temporal regularities of
spoken Swedish. Papers from the Institute of Linguistics,
University of Stockholm, Publication 21.
Martin, J.G. (1970). On judging pauses in spontaneous speech. Journal
of Verbal Learning and Verbal Behavior 9, 75-8.
Peterson, G.E. and Lehiste, I. (1960). Duration of syllabic nuclei
in English. Journal of the Acoustical Society of America 32,
693-703.
Postal, P.M. (1974). On Raising: One Rule of English Grammar and
Its Theoretical Implications. Cambridge: MIT Press.
98
Reis, M. (1973). Is there a rule of subject-to-object raising in
German? In Papers from the Ninth Regional Meeting of the Chicago
Linguistic Society (C. Coru, T.C. Smith-Stark, and A. Weiser,
Eds.). Chicago: Chicago Linguistic Society.
Rosenbaum, P.S. (1967). The Grammar of English Predicate Cornlement
Constructions. Cambridge: MIT Press.
Ross, J.R. (1969). A proposed rule of tree-pruning. In Modern
Studies in English (D.A. Reibel and S.A. Schane, Eds.). Englewood
Cliffs, N.J.: Prentice-Hall. Pp. 288-99.
99
Footnotes
Syllable lengthening has also been found at the beginnings of
major phrase boundaries in Swedish speech (Lindblom and Rapp, 1973).
However, this effect is small compared with the effect at the ends
of major phrase boundaries in Swedish, and the phrase-initial effect
has not been found for English (Klatt, 1975). The possibility of
a phrase-initial or clause-initial lengthening cannot be ruled out
for English, but for the arguments presented in this paper it is
necessary to assume only that such effects, if they exist, are small
compared with the effects in clause-final position.
2
This rule was originally included under the rule It-Replacement
(Rosenbaum, 1967). The rule must be kept distinct from a similar
rule proposed to raise a subordinate subject into superordinate subject
position. The latter rule converts the underlying structure of sen-
tences like It appeared that John was sick into John appeared to be
sick, with John as the raised element.
3 Under an alternative possibility, (1') would contain a single surface
clause, with the sentence node dominating be at breakfast deleted by
the convention of S-node Pruning (Ross, 1969; Hankamer, 1971; Reis,
1973). However, it appears that a Raising analysis, which preserves
a two-clause structure by selectively raising the subordinate subject,
is sufficient without pruning in English (see Reis, 1973). If,
conversely, S-node pruning is assumed, then the major result of
Experiment I to be discussed seems to have occurred at a level of
processing prior to the level at which pruning occurs.
100
An alternative to the Equi-NP analysis has recently been proposed
by Postal (1974). However, the location of the major clause
boundary under his analysis remains the same as in the Equi analysis.
5Although this analysis is assumed throughout the rest of the text,
the linguistic evidence favoring at least one aspect of the analysis
is less than compelling. John Robert Ross (personal communication)
has pointed out that while the underlying representation of (6)
may infact take an underlying complement (the critical assumption
for present purposes), the underlying complement verb may be a verb
of motion rather than to be. A main verb like want, used in Experiment III,
does appear to require an underlying to be in sentences like (6),
however. Ross cites as syntactic evidence of an underlying to be
for want complements the existence of idioms like I want headway made
on this by Friday and I want tabs kept on John. The argument applied
to headway goes as follows: (a) terms like headway can occur in under-
lying object position but not underlying subject position (*Headway is
fun); (b) in order to derive a sentence like I want headway made on this
by Friday, it is necessary to postulate a rule moving headway to the
left of make; (c) the relevant rule, Passive, requires to be; (d) thus,
the structure of I want headway made on this by Friday must include
an underlying to be which is present when Passive applies. For my
speech, the same argument applies to the main verb expect, lending
support for an underlying to be for the complement of this main verb
also.
6Klatt (1975) has noted that the modifications of vowel length are
more pronounced for vowels with inherently long durations.
101
7Speakers were not, however, asked to close their eyes while they
spoke.
8Although speakers did not place contrastive stress on any syllables
in the sentences, some difference in stress pattern existed between
the simple sentences (c) and (e) and the other sentences because of
the presence of the adjective big in the simple sentences. This ad-
jective was inserted to control for total number of words across
sentences. An account of the present results in terms of rhythmic
groupings of stressed syllables, however, would run counter to the
stress timing prediction needed to account for the results of the
second experiment in Chapter 2 (see especially Chapter 2, footnote 9).
Thus, in the absence of a more refined theory of stress timing,
the syntactic account to be proposed here should be favored.
9This and similar locations in the waveform differed to some extent
depending on the recorded amplitude of the speaker's voice. An
effort was made to keep the overall amplitude the same across speakers,
but in cases where the amplitude was perceptibly lower for a given
speaker, the amplitude display of each of the speaker's waveforms
was doubled to facilitate marking of the segment boundaries. Since
all comparisons of interest were across sentences within speakers, the
doubling of amplitude for a few speakers allowed for a more reliable
segment measuring procedure without introducing experimenter bias.
10 Recall that Chomsky predicts a difference in clause boundaries between
the infinitival complements of expect and persuade, whereas Rosenbaum
and Postal do not (see Table 1, surface structure predictions).
Assuming that the typical duration of [swedad] is longer than that of
102
[spEktad] on purely phonetic grounds (an assumption that is plausible
but has not yet been tested), the finding of longer durations for
[spektod] must be accounted for on syntactic grounds and would provide
a strong point in favor of a performance analog of Chomsky's analysis,
since on his account [spectad] is clause-final, whereas [swedad] is
not. The difference in verb durations between [spEktad] and [swedbd]
for simple sentences could also be accounted for on syntactic grounds,
by phrase-final lengthening, according to both Chomsky and Rosenbaum-
Postal.
1 The only difference found for the noun Kate in Experiment I which is
not easily accounted for by the phonetic factor demonstrated in
Experiment III was the significant difference between EXPECT-THAT
and EXPECT-SIMPLE, for which the duration of Kate was 13.0 msec
longer before at than before would. In Experiment III, a difference
for skate before [a] vs. [w] (at vs. with) in the same direction was
observed, averaging 8.0 msec, but this difference failed to reach
statistical significance with the relatively small N = 10.
12 Since a major clause boundary existed in the complement sentences
of Expect-type 2, an account of the single-clause lengthening for
believed and understood must include some reference to a non-syntactic
factor (e.g., focus) whose effectiveness overrides the syntactic
clause effect. Although some study of the relation between sentence
stress and focus has been conducted within a transformational frame-
work (see Akmajian and Jackendoff, 1970; Bresnan, 1971), work on
this general topic has not proceeded far enough to provide any good
hints about the precise account of the present cases of single-
clause lengthening.
103
1 3Clark (1973) has suggested the use of variance statistics like
mm F', where min F' (i,j) =F2 ( + p2), for testing the
generality of an effect across different sentence materials and dif-
ferent speakers using a design similar to that of Experiment III.
However, because of the lack of a large data base in this experiment,
it was decided to limit the analysis to computing the average dura-
tions across speakers for the sentences of each verb.
104
Table I
Syntactic Clause Boundaries Predicted on the Basis of
Two Linguistic Analysesa
Linguistic Analysis Complement Type
Rosenbaum-Postal
a. underlying structure
b. surface structure
expect to
expectA
Kate
expect that
expect^
expect A
persuade to
Kate A
Kate
Chomsky
a. underlying structure
b. surface structure
expectA
expectA
expect A
expect
KateA
Kate r
aBased on the sentences The host expected Kate to be at breakfast,
The host expected that Kate would be at breakfast, and The host persuaded
Kate to be at breakfast.
105
Table II
Mean Durations of the Last Two Syllables of
the Main Verbs in Experiment I
Sentence
Subject (a) (b) (c) (d) (e)
I.B. 407.6 402.4 396.2 391.8 392.4
R.B. 326.5 317.6 310.2 305.3 309.0
L.C. 376.2 363.2 351.7 371.6 328.3
B.F. 415.6 381.2 367.8 380.1 432.7
P.F. 358.3 335.9 330.9 342.2 325.8
R.F. 378.7 399.0 368.5 349.0 325.5
J.G. 381.3 384.5 363.5 354.8 328.1
R.G. 422.3 435.0 411.7 349.7 358.5
L.H. 366.3 365.6 366.9 368.8 343.1
M.J. 357.9 378.0 341.3 349.2 323.1
S.K. 353.5 353.5 355.5 314.3 364.2
B.P. 350.4 349.8 352.8 345.0 350.5
J.P. 369.7 362.4 370.9 357.8 350.1
R.T. 339.6 317.3 335.1 316.7 334.8
P.V. 342.9 348.9 323.8 326.5 306.9
Grand Mean 369.8 366.3 356.5 348.2 344.9
(a) = EXPECT-INF
(b) = EXPECT-THAT
(c) = EXPECT-SIMPLE
(d) = PERSUADE-INF
(e) = PERSUADE-SIMPLE
106
Table III
Mean Durations for the noun Kate in Experiment I
Sentence
Subject (a) (b) (c) (d) (e)
I.B. 194.5 210.9 208.4 192.7 190.4
R.B. 169.7 163.1 181.5 161.8 167.3
L.C. 171.9 184.8 178.6 159.2 187.3
B.F. 196.0 214.0 227.6 211.0 261.4
P.F. 212.4 211.9 228.1 223.2 244.5
R.F. 156.8 149.2 214.3 160.5 179.8
J.G. 168.6 164.3 200.5 175.3 184.7
R.G. 177.2 194.9 194.0 187.5 223.1
L.H. 168.7 167.5 172.4 162.1 158.2
M.J. 191.9 187.9 176.2 180.8 179.1
S.K. 172.4 174.9 194.7 182.6 209.3
B.P. 161.0 164.1 197.0 172.8 210.9
J.P. 157.7 163.9 153.1 161.6 147.9
R.T. 143.3 149.7 172.5 143.1 167.7
P.V. 167.2 172.8 170.9 172.9 177.5
Grand Mean 174.0 178.3 191.3 176.5 192.6
(a) = EXPECT-INF
(b) = EXPECT-THAT
(c) = EXPECT-SIMPLE
(d) = PERSUADE-INF
(e) = PERSUADE-SIMPLE
107
Table IV
Mean Durations of the Pronoun we and the Verb skate in Experiment II
Sentence Sentence
(a) (b) (c) (a) (b) (c)
Subject we skate
A.B. 165.1 175.2 162.6 391.0 419.1 401.9
D.B. 99.6 57.2 63.9 283.9 301.1 324.4
I.B. 127.6 135.8 129.1 366.1 384.5 403.9
J.B. 127.5 96.3 91.1 290.3 308.2 281.3
M.I. 61.9 66.5 70.7 254.9 267.1 274.1
C.K. 99.1 108.0 126.8 361.7 365.6 363.3
J.L. 59.7 96.7 83.0 282.2 293.5 325.7
S.L. 98.6 91.3 61.3 295.5 316.1 313.9
C.P. 128.3 115.5 144.3 390.4 399.2 411.3
J.P. 77.5 77.2 77.5 395.1 419.3 454.6
Grand Mean 104.5 102.0 101.0 331.1 347.4 355.4
(a) = We skate to the farm.
(b) = We skate with the crowd.
(c) = We skate at the pond.
108
Figure Caption
Figure 1. Hierarchical tree diagram representing the underlying
structure of sentences (1) [upper tree] and (2) [lower
tree], according to the analysis of complements provided
by Rosenbaum (1967).
SNP VP
Det N
t hs
the host
y
expect
NP
I
S
NP VP
kate V
I
beFP
NP
rep N
I I
at breakfast
S
NP VP
Det N
the
V NP S
I 1
persuade kate
NP VP
host
kate V NP
I /^\
be Prep N
I I
at breakfast
109
110
CHAPTER 4
Preposing Rules
ill
Preface to Chapter 4
"Naval historians think it worth while to argue about
Nelson's tactical plan at Trafalgar because he won the battle. It
is not worth while arguing about Villeneuve's plan. He did not succeed
in carrying it out, and therefore no one will ever know what it was."
[from R.G. Collingwood, An Autobiography, London: Oxford, 1939].
The above quotation applies by analogy to most scientific
work--reports and theoretical discussions center around the outcomes
of successful experiments, where "successful" is typically defined in
terms of data which permit the rejection of a statistical null
hypothesis. But null results may be informative too, particularly in
cases where (a) the likelihood of a Type II statistical error (by
which the null hypothesis is incorrectly accepted) is small, and
(b) the null result can be contrasted with positive results obtained
with the same experimental method.1 Because these conditions appear
to hold for some of the results reported in this chapter, it was
decided to include the report as part of this thesis. The report
should be of interest to anyone who wishes to pursue work using the
present method, as well as for those who would simply have liked to
know what Villeneuve's plan really was.
'When these conditions are satisfied, I take it that the null result
is "valid", in the sense used in the M.I.T. Graduate School Manual
(p. 40), and is therefore not merely desirable to report by required
under M.I.T.'s thesis regulations.
112
Linguistic Background
In addition to putative transformations like Raising, there
exist a number of other transformations which move NP or other
constituents.
Among these are the rules Passive, Topicalization, Adverb
Preposing, and Prepositional Phrase Preposing (hereafter, PP Preposing).
This chapter is a report of experiments designed to test the possibility
that the application of such movement rules influences speech timing,
either directly or indirectly via the particular form of surface
structure they produce.
We begin with the Passive transformation, a rule which has
received much attention in linguistic work. This rule has undergone
a number of revisions since its earliest formulation in generative
grammar by Chomsky (1957) .
Chomsky originally argued that sentences like (1) and (2)
should be related transformationally by an optionally-applied rule.
(1) John ate the apple.
(2) The apple was eaten by John.
According to this formulation of the rule, Passive operated to move
the underlying NP John to the right, attaching it to a ba-phrase, and
also moved the underlying NP the apple to the left, to the position
originally occupied by John.
Later, however, Chomsky reformulated the Passive rule as an
obligatory transformation which operated on structures like (3):
(3) NP-Aux-V-...-NP---by passive (where passive is a dummy
element in deep structure (Chomsky, 1965: 104).
113
Under this formulation, the transformation substituted the first NP
for the dummy element passive and placed the second NP in the original
position of the first.
The reformulation was designed to account for a number of
facts not handled well under the earlier optional formulation, including
the fact that passivization is normally restricted to sentences whose
verbs take manner adverbials (e.g., "with enthusiasm", "slowly").
Passive is blocked for a verb like resemble, for example (*Mary is
resembled by Jane), a verb which cannot take manner adverbials (*Mary
resembled Jane enthusiastically). By adding a base rule to the grammar
like (4), combined with the formulation of Passive in (3), the re-
striction of Passive to verbs that can take manner adverbials was
(4) Manner-+ by passive (Chomsky, 1965: 104).
captured by the grammar. In addition, the new formulation enabled
Chomsky to provide a somewhat more natural account of the existence
of "pseudo"-passives (e.g., The issues were enthusiastically argued
against).
Still further revision in the Passive rule was made when
Chomsky (1970) discussed this rule in relation to certain nominaliza-
tions under the lexicalist hypothesis. He noted that the two opera-
tions which comprise Passive, namely NP Postposing of the underlying
subject and NP Preposing of the underlying object, could be viewed
as independently applicable operations, assuming a relation between
passive and certain nominalizations whose existence could be accounted
for by the application of one or the other (but not both) of these
two operations (Chomsky, 1970: 203-204).
114
A final revision of Passive has been made which relies on the
recent "trace" theory of movement rules (see Chomsky, 1973, to appear;
Fiengo, 1974). Under this formulation, Passive and other NP movement
rules are assumed to leave a trace of their application. The trace
is morphologically null and acts like a bound variable, in the same
sense as an anaphoric element (Fiengo, 1974). The existence of such
morphologically null traces allows some generalizations to be stated
in a manner that appears to be more accurate and elegant than that
of any currently competing alternatives.
In the case of Passive, according to the new formulation,
a trace t is left behind in the place of both the underlying subject
and object after each is moved. However, the trace left by postposing
the underlying subject is erased by preposing the underlying object
to the position formerly held by this trace. Hence, only one trace
appears in the derived surface structure, occupying the position of
the underlying object, as shown in the string below:
(5) the apple was eaten t by John
An apparent advantage of the trace formulation of Passive is that it
allows the statement of a constraint on why idiomatic phrases like
kick the bucket can occur in only a non-idiomatic reading when they
are passivized (compare (6) and (7)):
(6) John kicked the bucket.
(7) The bucket was kicked by John.
Doubtless there exist other ways to state the idiomatic
restriction, but the trace theory as currently being developed appears
to provide a good way of stating this constraint as well as many others,
115
and the theory appears to be a highly promising approach to an account
of movement transformations.1
Objective
The goal of this preliminary study was to determine the timing
of words in a passive sentence, in comparison with a control sentence
containing a by-prepositional phrase. Because no previous studies of
timing have been concerned with passives, it was decided to provide
date on a number of candidate hypotheses, in particular, the three
listed below:
(a) the trace left by passive, according to Chomsky's most
recent formulation of movement rules, is accompanied by
a durational change in the immediately preceding or
following word--this hypothesis states that the trace
is accompanied by a phonetically observable change in
segment duration, even though the trace itself is
morphologically and phonologically null.
(b) the duration of a noun moved by Passive differs from the
duration of a corresponding stationary noun in the control
sentence
(c) the duration of the by in Passive differs from the duration
of the prepositional by
Because this study represented the first attempt to test timing
relations in passive sentences, the above hypotheses and the experimental
design itself were intentionally somewhat scattershot. If either
Hypothesis (a) or (b) turned out to be supported, then a rather profound
116
revision of current thinking about the control of speech timing would
seem to be in order. And from the linguist's vantage point, any
evidence favoring Hypothesis (a) would constitute a performance
analog to trace theory of a kind that could be very useful in guiding
further development of the theory itself.
Experiment I
Method
Subjects
Sixteen M.I.T. students and employees served voluntarily in
this experiment. All subjects were native speakers of English and
had no history of speech or hearing impariment.
Sentence Materials
Two sentences were used in the experiment:
(a) The students were seated by the widow. (Passive b)
(b) The students were seated by the window. (Prep. yi)
The sentences were matched for total number of words and syllables,
typical stress contour, and, indeed, differed from each other phonetically
only by the presence or absence of a nasal [n] in the first syllable
of the last word of the sentence.
The surface structure bracketing of both sentences was as
follows:
[SNPthe students]V[[AUXwere][Vseated] [NPby the wi(n)dow]]]
According to the trace theory, the passive version (Sentence (a)) also
contains a morphologically null trace preceding by.
Sentences (a) and (b) can also be said to differ in other
grammatical ways. The first NP of (a) is a patient whereas the first
NP of (b) is agentive; the verb in (a) is transitive whereas the
verb in (b) is intransitive; and the last NP in (a) is agentive
whereas the last NP in (b) is a locative prepositional phrase.
Procedure
The general testing procedure described in Chapter 1 was used
here. Speakers were told at the outset of the experiment to read
Sentence (a) as a passive sentence rather than on its possible reading
as a sentence containing a locative prepositional phrase (e.g., The
students were seated by (=beside) the widow.) Each sentence was
practiced individually before it was read for recording purposes 6
times in succession. Speakers began with Sentence (a) or (b) according
to nndom assignment. The first 5 occurrences of each sentence
(excluding mispronunciations) were digitized at 10 kHz and analyzed
for the durations of each of the first 5 words of the sentence. The
measurements of each word included all phonetic segments of the word,
with the sole exception that any pre-voicing for by was excluded from
the measurement of this word (in this case, pre-voicing was noticed
in a few utterances, distinguishable from the voicing of the previous
word by a change in the amplitude of voicing). The occasional pre-
voicing for by was excluded to minimize the duration variability of
this word.
118
Results and Discussion
The mean duration of each of the 5 words for each speaker
is shown in Table I. Two-tailed t-tests for matched pairs were
Insert Table I about here
applied to the mean durations for the group of speakers and revealed
no statistically significant differences between the two test condi-
tions for each word (p > .20 in each case). The absence of any
significant effects indicates that the data provide no support for
any of the three hypotheses outlined in the Objective. Thus, there
is no evidence to support the notion that morphologically null traces
are accompanied by a phonetic change in duration, that a moved NP
differs in duration from a stationary NP (in terms of a sentence deriva-
tion), or that the duration of passive by differs significantly from
prepositional by.
Of the 16 speakers, two (J.L. and P.V.) produced an average
lengthening of all 5 words in the passive sentence in comparison with
the prepositional sentence, while one speaker (A.B.) produced an
equally systematic effect for all 5 words in the opposite direction.
It is quite possible that these individual trends represent changes in
overall speaking rate rather than changes in duration as a direct
function of the grammatical structure. Further work is required,
however, to test the latter possibility.
119
Other Preposing Rules
A second experiment was designed to provide another test of
a phonetic manifestation for morphologically null traces (Chomsky,
1973, to appear). Three preposing rules were studied--Topicalization,
Adverb Preyosing, and PP Preposing.
The first of these rules, Topicalization, serves to move an
NP to the beginning of its clause, transforming the structure of
sentences like (8) into (9):
(8) We taped the concert with the Ampex recorder.
(9) The concert we taped with the Ampex recorder.
This construction rarely appears in written text, but it occurs
commonly in normal conversation. In my speech, the moved NP is
lengthened, and a noticeable pause occurs immediately afterwards. The
concern here, however, is not with any lengthening that may occur for
the moved NP but with any change in duration that may occur just
prior to its underlying position, where a morphologically null trace
appears in surface structure. This trace occurs immediately after
the verb (Fiengo, 1974).
The second transformation, Adverb Preposing, moves an adverb
to the beginning of its clause, accounting for the relation between
(10) and (11):
(10) We taped adeptly with the Ampex recorder.
(11) Adeptly we taped with the Ampex recorder.
Here, as in the case of Topicalization, the application of the preposing
rule results in a construction that rarely appears in written English.
In my speech, Sentence (11) is accompanied by lengthening of the
120
preposed adverb and a puase between this word and the rest of the
sentence, similar to the case for topicalized sentences. However,
unlike Topicalization, there is no currently available lingristic
basis for postulating a trace at the site of adverb movement. Trace
theory, as developed so far, pertains only to rules that move NPs.
A third preposing rule, PP Preposing, moves a prepositional
phrase to the beginning of its clause, deriving the structure of
(13) from (12):
(12) We taped with the Ampex recorder at sundown.
(13) At sundown we taped with the Ampex recorder.
This construction appears more commonly in written and spoken English
than either of the two preceding preposed constructions. Like the
others, lengthening of the moved NP is noticeable inmy speech, as is
the presence of a pause following this NP. But unlike the rule of
Topicalization, the location of the trace left by PP Preposing is
believed to occur after the direct object NP, not immediately after
the verb. Thus, according to the hypothesis that traces are ac-
companied by a change in the duration of an immediately preceding
word, the durations of the verb in topicalized sentences should differ
from the verb durations in control sentences (no preposing), whereas
such a difference should not be found for the duration of the verb
in a PP-preposed sentence compared with its control.
121
Experiment II
Method
Subjects
Nineteen M.I.T. students and employees served as volunteers
in the experiment, having the same qualifications as the subjects
who served in Experiment I.
Sentence Materials
Five sentences were used in the experiment:
(a) I sang before we taped with the Ampex recorder. CONTROL
(b) At lectures we taped with the Ampex recorder. PP PREPOSING
(LOCATIVE)
(c) At sundown we taped with the Ampex recorder. PP PREPOSING
(TEMPORAL)
(d) Adeptly we taped with the Ampex recorder. (ADVERB PREPOSING)
(e) The concert we taped with the Ampex recorder. (TOPICALIZATION)
Each sentence contained from 7 to 9 words and from 12 to 13 syllables.
Each sentence contained the same approximate stress contour in the
vicinity of the key word.
Procedure
The general testing procedure described in Chapter 1 was
used here. Each speaker read each sentence 6 times, beginning with
(a) or (e) according to random assignment. The duration of the key
word tape was measured. The measured segment included the [t]
release burst and the vowel [e], but did not include the [p] closure
122
interval or any [p] release burst. The decision not to include the
latter segments was made to minimize the duration variability.
Results and Discussion
The results appear in Table II. The only consistent trend
emerging from the data is for the verb of both locative and temporal
PP preposing sentences (b) and (c). The verbs in both sentences were
shorter in duration than the verb duration of the control sentence (a).
Insert Table II about here
The results of the experiment provide no support whatsoever
for the hypothesis that morphologically null traces are accompanied
by a change in the duration of a preceding word. According to this
hypothesis, a comparison between the verb durations of the preposed
and control sentences should have revealed consistent differences
for Topicalization only, yet the only difference observed was found
for PP Preposing, where no trace is postulated immediately after
the verb.
These results, taken together with the null results of
Experiment I on passives, suggest that, to the extent that morphologically
null traces are processed during sentence production, such traces are
accompanied by no changes in the duration of an immediately preceding
word. The data from Experiment I suggest that no changes in the
duration of an immediately following word occur either.
123
This particular set of tests was conducted primarily in
search of support for an a priori very improbable, but theoretically
exciting, hypothesis concerning a phonetic manifestation of traces.
No support was found for this hypothesis, but further work using
preposed constructions may still be of value toward developing a
theory of speech timing. In particular, it was noted above that,
for my speech, preposed constituents appear to be lengthened and
accompanied by a pause in comparison with non-preposed constructions.
Although duration measurements were not made forpreposed constituents
in the two experiments, I noticed the same trend as found in my
speech while listening to the utterances of nearly all of the 19
speakers. On the basis of such listening, it appears that lengthening
of the final word of a preposed constituent in the cases of Topicaliza-
tion, Adverb Preposing, and PP Preposing averages at least 25 msec,
and perhaps as much as 75 msec. If acoustic measurements verify these
estimates, then a fairly interesting question can be raised, namely,
why is lengthening so pronounced for these preposing rules when it is
altogether absent in the case of passivization (cf. Experiment I)?
One possibility is that Passive, unlike the other rules, incorporates
both NP Postposing and NP Preposing, such that the surface structure
contains NPs in roughly the same positions as they appear in under-
lying structure. The other preposing rules produce a more drastic
alteration of the normal sentence structure. In linguistic terminology,
this impression has been given a formal representation by Emonds
(1970). Emonds defines two types of transformational rules, root
transfromations and structure-preserving transformations. Root
transformations move constituents into positions which could not be
124
generatea by the phrase structure rules of the base component of
grammar, while structure-preserving transformations may move constituents
only into positions which could be so generated. According to this
formulation, Topicalization, Adverb Preposing, and PP Presposing are
all root transformations, whereas Passive is structure-preserving.
It is possible that this difference in transformational type is
associated with a major difference in the control of speech timing;
i.e., that root transformations but not structure-preserving trans-
formations produce changes in word duration. Work directed at testing
this principle may lead to a firmer test of whether speech timing is
controlled by a level of processing comparable to the transformational
component of generative grammar.
125
References
Chomsky, N. (1957). Syntactic Structures. The Hague: Mouton.
Chomsky, N. (1965). Aspects of the Theory of Syntax. Cambridge:
MIT Press.
Chomsky, N. (1973). Conditions on transformations. In S.R. Anderson
and P. Kiparsky (Eds.) A Festschrift for Morris Halle. New York:
Holt, Rinehart, and Winston. Pp. 232-286.
Chomsky, N. (to appear). Trace theory. Linguistic Inquiry.
Emonds, J.E. (1970). Root and Structure-Preserving Transformations.
Unpublished Ph.D. Dissertation, M.I.T., Cambridge, Ma.
Fiengo, R. (1974). Semantic Conditions on Surface Structure. Unpublished
Ph.D. Dissertation, M.I.T., Cambridge, Ma.
126
Footnote
In Chomsky's most recent formulation (class lectures, 1975), the
application of trace theory may result in an elimination of the Passive
rule as such, replacing it and other NP movement rules by a general
rule Move NP, whose application is contrained by metaconditions on
transformational theory.
127
Table I
Mean Word Durations of Each Speaker for Sentences (a) and (b)
of Experiment I
Sentence
Sentence
(a):
(b):
The students were seated by the widow.
The students were seated by the window.
Word
Sentence
Speaker
A.B.
L.C.
B.F.
P.F.
R.F.
J.G.
R.G.
M.I.
C.K.
S.K.
J.L.
S.L.
E.M.
J.P.
R.T.
P.V.
Grand Mean
the students
(a) (b) (a) (b)
89.4
80.4
66.1
45.2
64.5
55.2
64.7
51.5
70.5
49.6
54.5
51.2
62.1
63.9
65.5
41.7
61.0
89.8
65.3
66.3
59.9
61.3
45.8
70.6
56.5
67.9
48.8
49.9
51.8
69.0
69.1
62.9
38.7
60.9
519.9
383.7
577.3
492.3
424.4
414.5
525.1
408.8
471.9
387.5
496.7
454.8
435.1
451.5
497.0
454.4
462.2
565.7
394.9
564.2
404.9
407.6
445.6
542.5
377.9
466.0
418.0
468.9
463.4
423.1
425.2
497.1
438.0
456.4
were seated
(a) (b) (a) (b)
111.2
108.6
105.6
114.1
109.7
109.4
114.9
83.4
84.8
115.6
109.4
101.8
96.9
78.2
127.6
141.9
107.1
141.5
128.4
109.1
110.4
118.9
107.1
143.9
69.1
81.8
118.8
103.3
112.5
89.7
84.7
132.9
124.8
111.1
448.9
346.4
384.3
382.4
362.0
355.0
403.1
335.9
376.3
383.4
462.0
377.9
344.7
376.1
331.0
371.2
377.5
487.7
358.6
417.0
391.6
308.7
347.3
397.1
322.4
394.7
409.4
445.8
371.5
335.4
368.6
332.6
361.3
378.1
by
(a) (b)
199.7
155.9
160.8
168.5
154.0
133.3
156.7
145.2
147.3
152.4
255.7
152.8
167.3
136.7
153.9
131.9
160.8
233.0
167.4
161.3
171.5
137.4
138.3
155.4
142.0
168.0
171.6
235.4
156.1
164.4
148.7
134.7
120.9
162.9
128
Table II
Mean Verb Durations of Each Speaker for Sentences (a)-(e)
of Experiment II
Sentence
Speaker (a) (b) (c) (d) (e)
A.B. 209.2 205.8 209.9 214.5 227.8
A.Br. 228.0 228.1 215.1 211.7 195.5
D.B. 193.9 196.7 197.5 196.0 200.8
I.B. 184.6 178.7 184.0 189.4 192.8
J.B. 184.6 172.5 175.2 156.5 169.1
S.B. 184.1 172.7 175.8 187.2 176.4
D.D. 137.2 128.7 129.3 149.2 140.7
J.D. 177.9 179.4 184.4 175.7 177.2
L.D. 177.6 179.0 181.3 169.8 175.5
M.I. 184.1 170.7 161.2 174.5 170.3
C.K. 175.5 181.5 176.7 196.7 186.4
D.K. 191.4 188.1 190.8 184.8 180.9
J.L. 197.1 184.8 188.8 205.6 198.1
S.L. 181.4 168.4 169.0 162.2 158.9
S.Lo. 145.9 152.4 150.8 155.0 163.6
E.M. 176.6 160.0 170.9 165.3 158.8
E.Mc. 178.4 185.3 183.9 184.5 200.9
C.P. 184.5 188.2 171.8 197.6 188.5
J.P. 193.4 194.0 169.4 184.2 190.8
Grand Mean 183.4 179.7 178.2 182.1 181.7
Sentence (a): I sang before we taped with the Ampex recorder.
(b): At lectures we taped with the Ampex recorder.
(c): At sundown we taped with the Ampex recorder.
(d): Adeptly we taped with the Ampex recorder.
(e): The concert we taped with the Ampex recorder.
129
CHAPTER 5
Speech Timing of Coreference
130
Abstract
Speakers lengthened the duration of a noun when it was referred
to again later in the same sentence by alpronoun. The effect was
observed for the very first word of an utterance and for coreferents
that spanned a major syntactic clause boundary.
131
CONSIDER the following ambiguous sentence:
(1) Kate told Jane that she didn't need to pass French.
This sentence contains two clauses, whose underlying representations,
according to generative grammar, are roughly [Kate told Jane X1 and
[Kate (or Jane) didn't need to pass French.] The second clause is
subordinate to the main clause.
The pronoun she in the subordinate clause can refer back to
either Kate or Jane. Linguists have noted that a speaker may place
emphatic stress on a pronoun in ambiguous sentences like (1) in case
the speakre intends the pronoun to be a coreferent of the more distant
antecedent (e.g., Kate).1,2 Since duration is a major concomitant
of stress, it was expected that in sentences like (1) the duration
of she would be much longer when this pronoun refers back to Kate.
The present experiment shows, however, that speakers more con-
sistently lengthen the duration of the antecedent noun itself, by a
small amount averaging about 8 msec. The result indicates that a
semantic level of processing partly controls the timing of words in
speech. Furthermore, the result suggests that the timing of coreferents
may take place at a different stage of processing than one limited
to the domain of a single clause.4
The eAperiment included 12 speakers from the M.I.T. community.
Each speaker was presented Sentence (1) and was instructed to consider
each of its two possible meanings. The speaker was instructed to
practice saying the sentence according to each meaning and to try
to make the intended meaning understandable to a listener. Following
practice without feedback from the experimenter, each speaker uttered
each of the two versions of Sentence (1) 6 times in succession for
132
purpose of recording, beginning with either version according to
random assignment. The waveforms of the first 5 occurrences of each
sentence for each speaker were digitized at a sampling rate of 10 kHz
and measured for the duration of the words Kate, Jane, and she.
The measurements were made from digitized oscillographic traces of
the speech waveforms with the aid of a computer controlled cursor.5
The offset of visible waveform periodicity was taken as the end of
each word segment; hence, the [t] closure interval for Kate or any
following [t] release burst was not included as part of the duration
of Kate. The decision not to include such segements was made in order
to minimize duration variability. The reliability of each measurement
was estimated to be within + 2 msec for Kate and + 3 msec for both
Jane and she. The measurements for Jane for one speaker were excluded
from the data analysis because the reliability of the measurements
exceeded + 3 msec.
The results for the version of Sentence (1) in which she
referred to Kate showed the following average durations for the group
of speakers: 182.8 msec for Kate, 318.5 msec for Jane, and 232.1 msec
for she. For the version of Sentence (1) in which she referred
to Jane, the average durations were: 174.8 msec for Kate, 325.8 msec
for Jane, and 196.8 msec for she. The average duration of Kate was
thus 8.0 msec longer when she referred back to Kate, and the average
duration of Jane was 7.3 msec longer when she referred back to Jane.
This lengthening effect was obtained with 9 of the 12 speakers for
Kate and with 9 of 11 speakers for Jane. In addition, the average
duration of she was 35.3 msec longer when she referred back to Kate,
its more distant antecedent. This effect, however, was observed for
133
only 7 of the 12 speakers. For 4 of the speakers, the effect averaged
more than 65 msec and represented a clearly perceptible indication
of emphatic stress. In summary, while the duration of she for some
speakers was much longer when it referred to the more distant antecedent,
a smaller but more consistent effect of lengthening was obtained for
the duration of the antecedents themselves.
Antecedent lengthening may be controlled by the relation of
coreference per se or by a relation of focus, whereby the antecedent
is lengthened because it happens to serve as the main information-
bearing referent of the sentence. Regardless of how this distinction
is resolved,4 the present effect indicates that speakers time the
durations of words with reference to their semantic relation to other
words in a sentence. This semantic influence can be observed even
for the first word of an utterance. In addition, the influence
is not blocked by the presence of a major syntactic clause boundary.
The latter condition suggests that the level of processing at which
semantic control of timing takes place occurs prior to a level of
processing that is limited to clause-by-clause computation. 4 ,6
Further work should be directed at the question of whether
antecedent lengthening is observed under conditions that more closely
approximate a natural speaking situation. For example, antecedent
lengthening should be studied for non-ambiguous sentences otherwise
similar to (1) [e.g., Kate told Don that she didn't need to pass
French] in which the speaker is asked to read the sentences in
conversational context.
134
This work was supported by a National Science Foundation
Graduate Fellowship and by the National Institutes of Health. I
thank Dr. Merrill F. Garrett for advice.
135
References
1Akmajian, A. and Jackendoff, R.S., Ling. Incp, l, 124-126 (1970).
2Cantrall, W.R. in Papers from the parasession on functionalism,
(edit. by Grossman, R.E., San, L.J., and Vance, T.J.), 36-46
(Chicago Linguistic Society, Chicago, 1975).
3
Fry, D.B., J. Acoust. Soc. Amer., 27, 765-768 (1955).
4Cooper, W.E. Syntactic control of timing in speech production
(Ph.D. Dissertation, M.I.T., 1976).
5Huggins, A.W.F. Q. Prog. Rep. Res. Lab. Electr., M.I.T., 95, 81-83
(1969).
6Garrett, M.F. in Advances in learning theory and motivation (edit. by
Bower, G.) (Academic Press, New York, in the press).
136
CHAPTER 6
Speech Timing of Coreference. II. Precede and Command Relations
137
Introduction
This chapter includes the report of a second experiment on
coreference. The results provide a further demonstration that the
control of timing is not necessarily restricted to the domain of a
single clause.
The present study concerns two syntactic conditions on
pronominalization. One of these conditions, termed the precede
condition (Lanacker, 1969), involves the linear relation between a
pronoun and its antecedent. The other condition, termed command,
involves a hierarchical relation between two coreferents. By
comparing word durations in sentences in which one or the other of
these two conditions is violated, one can determine whether the linear
and hierarchical syntactic constraints on pronominalization play
different roles in the control of speech timing. In addition, the
comparison between these two conditions allows one to provide a further
test of whether coreference relations play a role in the control of
timing and whether such control, if it exists, is restricted to the
domain of a single clause (see the preceding chapter for discussion).
Langacker (1969) formulated a general constraint on pronominal-
ization as follows: a pronoun cannot both precede and command its
antecedent (the term antecedent as used by linguists refers to the
definite noun phrase that is coreferential with a pronoun--the ante-
cedent need not precede the pronoun). The hierarchical notion command
was defined by Langacker in terms of structural nodes A and B as follows:
A commands B if the sentence-node most immediately dominating A also
dominates B, and if A and B do not dominate each other. By applying
138
the definition of command to the example structure in Figure 1, we
note that A commands B, C, and D; B commands A, C, and D; C commands
only D; and D commands only C.
Insert Figure 1 about here
Langacker's formulation of the constraint on pronominalization
in terms of the notions precede and command was designed to account
for the ungrammaticality of sentences like (1) below, in which the
indexed nouns are intended to be coreferential:
(1) *He will leave if Heath. is in real trouble.
Ross (1969) and Postal (1970) also noted and discussed this constraint.1
Both of the normal precede and command relations between
pronoun and antecedent must be violated in order for pronominalization
to be blocked. If either of the relations is violated singly, fully
grammatical sentences are obtained, as in (2) and (3):
(2) If he. leaves the state, Heath. will be in real trouble.
(3) If Heath. leaves the state, he. will be in real trouble.
Sentence (2) contains a violation of the normal precede relation,
resulting in so-called backwards pronominalization. Sentence (3), on
the other hand, contains a violation of the normal command relation,
with the definite noun phrase Heath occurring in a conditional clause
dominated by the main clause containing the pronoun.
Sentences (2) and (3) were used here as materials in a sentence-
reading experiment, similar in design to the experiment reported in the
previous chapter. Instead of measuring the durations of the pronoun and
139
antecedent as in the earlier experiment, it was decided to measure
the duration of the key word state. This word was chosen because
there exists no conceivable difference in the semantic focus assigned
to this word in Sentence (2) vs. (3). Thus, any duration effects
obtained must be attributed either directly to a difference between
the coreference relations of precede and command or to another factor
that depends on this difference.
Experiment Method
Subjects
Twelve M.I.T. students and employees served voluntarily in
the experiment. All subjects were native speakers of English with
no history of speech or hearing impairment. All 12 subjects served
in the experiment described in the previous chapter during the same
test session.
Sentence Materials
Sentences (2) and (3) were used in the experiment. These
sentences are repeated below for convenience:
(2) If he leaves the state, Heath will be in real trouble.
(3) If Heath leaves the state, he will be in real trouble.
Each of the two sentences contained 11 words and 12 syllables.
The sentences differed from each other solely in the location of the
coreferential nouns Heath and he, whose positions were interchanged in
the two sentences. In (2), the pronoun preceded but did not command
its antecedent, while in (3), the pronoun commanded but did not precede
its antecedent. To control for any possible effects of phonetic
environment on the duration of the key word state, this word was bounded
140
on the left in both sentences by the same word, the, and on the
right by either Heath or he, containing the same initial phonetic
segments. The latter words were also separated from the key word
by a major syntactic clause boundary, making it unlikely that any
difference in the stress (with greater stress on Heath) or phonetic
structure of the two words systematically influenced the duration
of the key word (cf. Huggins, 1974).2
Procedure
The subjects were tested individually according to the general
protocol used in the experiment reported in the previous chapter. In
this experiment, however, the speaker was told not to place contrastive
or emphatic stress on any words in order to minimize variability and
avoid any systematic stress differences on the coreferential nouns
Heath and he that might influence the duration of the key word (see
Footnote 2). Each speaker was instructed to practice reading each of
the two sentences in a natural speaking voice and was encouraged to
read the sentences as unitary utterances rather than word-by-word
as in unpracticed reading. Following practice, each speaker read each
of the two sentences 6 times in succession. The speaker began with
Sentence (2) or (3) according to random assignment. The first 5
occurrences of each sentence were digitized at a sampling rate of
10 kHz and analyzed for the duration of the key word state. The
onset of this word was taken as the beginning of visible frication
in the digitized oscillographic trace. The offset of the word was
taken as the end of visible periodicity in the vowel. The closure
interval of the syllable-final [t] and any closure release burst was
not included as part of the word duration in order to minimize
141
duration variability. The reliability of each measurement was
estimated to be within + 3 msec.
Results and Discussion
The average durations of the word state in each sentence
for each speaker are presented in Table 1. The average duration for
the 12 speakers in Sentence (2) was 13.3 msec longer than in Sentence
(3). The longer duration in (2) was found for 9 of the 12 speakers,
Insert Table 1 about here
and the lengthening effect for the speakers as a set was statistically
significant (t = 2.636, p<.05, df = 11; two-tailed t-test for matched
pairs).
The present effect provides a further demonstration that
coreference relations play a direct or indirect role in the control
of speech timing and that this influence is not restricted to the
domain of a single clause. The only surface difference between the
two sentences in the experiment was the location of the two coreferents,
which were members of separate clauses in both sentences.
The experiment of the previous chapter also showed a relation
between coreference and speech timing. That experiiaent was conducted
with a structurally ambiguous sentence and speakers were instructed
to read each version of the sentence so as to make the intended meaning
142
understandable to a listener. The previous results could have
reflected a capability on the speaker's part to take account of
coreference relations in speech timing in a special circumstance
and not in a setting approximating normal speech. In the present
experiment, however, the sentence materials were not structurally
ambiguous, and speakers were simply instructed to read the sentences
in their normal speaking voice. Thus, the results indicate that
coreference relations play a role in timing for non-ambiguous as well
as ambiguous sentences and suggest that such relations may play a
role in the timing of natural speech.
In addition to these general considerations, it may be
possible to provide a specific account of the direction of effect
observed in the present experiment. One possibility is that the
duration of state was longer in Sentence (2) because violation of the
precede relation resulted in a structure that is less typical than
the structure in (3). At the word level, it has been shown by Coker,
Umeda, and Browman (1973) that the duration of atypical or low-frequency
words is longer than the duration of more typical words. If this
principle can be extended to the level of sentence structure, and if
violation of the precede relation indeed produced a less typical
structure than violation of the command relation, then an account would
be provided for the greater duration of the clause-final word in (2).
143
Footnotes
This work was supported by a National Science Foundation
Graduate Fellowship, a grant from the Sloan Foundation to the
Department of Psychology, M.I.T., and by NIH Grant HD-05168. I
thank Dumont Billings for assistance.
1Postal (1970: 20ff.) notes that Langacker's precise
formulation of the constraint may require modification. In addition,
Postal shows that this major constraint on proniminalization must be
accompanied by other constraints (in particular, a cross-over restric-
tion). For present purposes, however, it is sufficient to assume
that a hierarchical notion similar to Langacker's command is required
to account for the major constraint on pronominalization under
discussion.
2In order to test this possibility directly in future work, a
control experiment should be conducted using sentences like (W') and
(2'):
(1') If Joan leaves the state, Heath will be in real trouble.
(2') If Joan leaves the state, he will be in real trouble.
144
References
Coker, C.H., Umeda, N., and Browman, C.P. (1973) Automatic synthesis from
ordinary English text. IEEEAudio and Electroacoustics AU-21,
293-297.
Huggins, A.W.F. (1974) An effect of syntaz on syllable timing. Quarterly
Progress Report of the M.I.T. Research Laboratory of Electronics
114, 179-185.
Langacker, R.W. (1969) Pronominalization and the chain of command.
In (D.A. Reibel and S.A. Schane, Eds.) Modern Studies in English.
Prentice-Hall, Englewood Cliffs, N.J. Pp. 160-186.
Postal, P.M. (1970) Cross-Over Phenomena. Holt, Rinehart, and Winston,
New York.
Ross, J.R. (1969) On the cyclic nature of English pronominalization.
In (D.A. Reibel and S.A. Schane, Eds.) Modern Studies in English.
Prentice-Hall, Englewood Cliffs, N.J. Pp 187-200.
145
Table I
Mean Durations of the Key Word state for Each Speaker in Sentences (2) and (3)
Sentence (2)
303.7
328.5
354.4
341.4
348.5
381.3
348.3
325.3
367.7
391.0
360.8
435.7
Sentence (3)
306.6
337.8
348.8
326.9
319.8
394.3
322.6
310.8
341.8
361.5
339.2
416.8
Grand Mean 357.2 343.9
Speaker
R.B.
L.H.
M.I.
M.J.
C.K.
J.L.
S.L.
E.M.
B.P.
C.P.
J.P.
K.P.
0
146
Figure Caption
Figure 1. Structural respresentation to illustrate command relations.
147 
s 
A B 
s 
C D 
148
CHAPTER 7
Conclusion
149
In addition to the experiments described in preceding chapters,
a number of other experiments were conducted using the same testing
method, typically using only from 3 to 7 speakers in each experiment.
Because of the small data base for these experiments, few if any firm
conclusions can be drawn from the results. Current work is aimed at
building larger data bases for some of these experiments, and the final
results of this work will be reported elsewhere (Cooper, in preparation).
However, it seems useful to provide a preliminary sketch of these
additional studies here in order to give some indication about the
direction which the work is currently taking.
The experiments can be classified into four main areas, concerning
(1) deep structure ambiguities, (2) deletion rules, (3) movement rules,
and (4) clause types. The work in each of these areas is discussed
briefly below, followed by concluding remarks summarizing the direction
of the work as a whole.
1. Deep Structure Ambiguities
The experiments in the preceding chapters were primarily designed
to uncover possible effects controlled by transformational or surface
levels of syntactic processing. However, some preliminary work was also
aimed at the possibility that underlying syntactic structure plays a
role in the control of speech timing. To test the relevance of this
syntactic level of processing (if it indeed exists), the timing of words
in two ambiguous sentences was studied. These sentences, presented
below in (1) and (2), represent so-called "deep structure ambiguities"
because they have two possible underlying syntactic representations
but only a single surface structure.
150
(1) Which criminals have the police surrounded?1
(2) The shooting of the criminals was appalling.
Sentence (1) can question either which criminals surround the
police or which police surround the criminals. In essence, the
ambiguity concerns the underlying grammatical roles of subject and
object. Sentence (2) concerns a statement about either the criminals'
act of shooting or someone else's act of shooting the criminals.
The duration of each word in both Sentences (1) and (2) were
measured for a group of 5 speakers. At the outset of the test, each
speaker was told that the sentences were ambiguous and was informed
about the nature of the ambiguity. The speaker was then told to
practice saying each of the intended meanings of each sentence so
as to make the meaning understandable to a listener if possible.
Each version of each sentence was then recorded 6 times.
The word durations showed some sizable differences between
the two meanings for individual speakers, in one case averaging more
than 100 msec. However, these effects were not very systematic in
direction across different speakers, and a larger data base is clearly
needed to test for population effects. The presence of large and
consistent effects within the data from each subject, however, provided
rather convincing support for the notion that deep structure differences
can influence speech timing. It should be noted, however, that the
evidence provided here is also compatible with a model of timing in
which logical relations control timing directly rather than via an
underlying syntactic representation.
151
2. Deletion Rules
Additional work on deletion rules has centered on the putative
rules of to be deletion (Borkin, 1973) and will deletion (Vetter, 1973).
The work on to be deletion, designed to test for evidence of under-
lying and/or transformational structure effects on timing, involved
the two sentences below.
(3) John appeared abruptly at the meeting.
(4) John appeared abrupt at the meeting.
While both (3) and (4) contain a single clause in surface structure,
(4) contains a putative two-clause underlying structure, represented
roughly by the strings John appeared to be X and John was abrupt at
the meeting. If the underlying clause boundary occurring after the
verb appeared plays a role in word lengthening, then it would be
expected that the duration of this verb should be longer in (4) than
in (3). This hypothesis was tested with 6 speakers, one of whom showed
the predicted effect by an average of 39.6 msec. However, only two
other speakers of the group showed a trend in the same direction,
indicating that a larger data base is required before any firm conclusions
cam be reached about the role of underlying clause structure for this
sentence construction. Work on a putative rule of will deletion
involved the following sentences:
(5) Tomorrow I will face the students cheerfully.
(6) Next week I face the committee on academic standing.
(7) I will face the committee tomorrow.
(8) On Tuesday John will face the new students.
(9) Next month I face the prospect of hard times.
152
(10) I will face the dean on Monday morning.
(11) On Tuesday John will face the new students.
(12) Next week I will face the president's envoy.
(13) Tom will face the crowd on Thursday.
These sentences were included as part of a 39-sentence list used in
the earliest stages of pilot work. Notice that, in sentences (6) and
(9), future tense is indicated by the presence of appropriate lexical
information, but a future tense marker will is absent, unlike the
remaining sentences in the list. No durational effect of this
deletion was found for the key word face in these sentences in data
for three speakers. In addition, no consistent differences in the
duration of the key word were obtained depending on whether the temporal
prepositional phrase in the sentences occurred in clause-final position
or in a preposed position (see Chapter 4) (compare, for example,
Sentences (5) and (7)).
The work on to be deletion is currently being extended because
of its importance in testing for effects of underlying clause structure.
The work on will deletion is not being continued at present because no
major theoretical issues appear to depend on the outcome of such work.
3. Movement Rules
Additional studies of movement rules have included sentences
involving A-Verb Raising (Postal, 1974) and Particle Movement. The
former rule converts the structure of a sentence like (14) into (15),
while the latter rule converts the struct of (16) into (17) (or 17
into (16)--the direction of movement being unclear in the case of this
153
rule--see Jacendoff, 1975 for discussion favoring the derivation of
(16) from (17)).
(14) It seems that John is angry.
(15) John seems to be angry.
(16) John phoned up Martha.
(17) John phoned Martha up.
The preliminary results obtained with these constructions
indicate that some speakers produce systematic differences for the
sentences of each of the two types. However, no firm conclusions
can be reached concerning the presence or absence of population
effects.
4. Clause Types
Of major interest in current timing research are differences
in the magnitude of clause-final lengthening obtained with different
types of clauses, e.g. complement, relative, conditional, causal.
It was pointed out earlier that audible differences in magnitude exist
in the lengthening produced by different clause types, yet no system-
atic account of these differences or of their theoretical basis exists.
To provide information on the empirical issue, a test using 6 speakers
was conducted with sentences containing two clauses, one main clause
and one subordinate clause, the -latter being either a conditional
clause (introduced by if or unless) or a temporal clause (introduced
by before or after). The following four sentences were used:
(18) If we reveal Kate's hopes the Doctor will see her.
(19) Unless we reveal Kate's hopes the Doctor will see her.
(20) Before we reveal Kate's hopes the Doctor will see her.
154
(21) After we reveal Kate's hopes the Doctor will see her.
The duration of the vowel Eo] of the key word hopes was measured for
duration in each sentence. The average durations for the key word
of the two conditional clauses (18) and (19) were 151.6 msec and
159.8 msec, respectively, while the average durations for the key
word of the two temporal clauses (20) and (21) were 147.0 msec and
147.5 msec, respectively. On the basis of these preliminary data,
it appears that clause-final lengthening is more pronounced for
conditional than for temporal subordinate clauses when followed by
a main clause. However, this difference is small compared to the
differences expected to exist amoung some other clause types. For
example, it is expected, based on impressions of listening to
conversational speech, that clause-final lengthening at the end of
a complement clause is very small in magnitude compared with the
lengthening effects observed at the ends of other subordinate clauses,
as well as at the end of a conjoined clause. Further experiments are
required to verify and quantify this impression. As noted in the
introduction, such differences in magnitude should probably be imple-
mented in programs to synthesize speech by rule. A theoretical account
of the differences may rely on the degree to which two clauses separated
by a clause boundary are semantically related to one another (the closer
the relation, the less lengthening). However, the latter possibility
requires a much more precise definition.
Another area of current study on clause types for which some
preliminary data have been obtained concerns a comparison between the
timing of a word preceding a complement vs. relative clause.2 Four
155
sentences were used in this experiment:
(22) I regretted the fact that John killed Sam.
(23) I regretted the fact that John reported.
(24) I regretted the fact that John died.
(25) I regretted the fact that John announced.
Sentences (22) and (24) contain a main clause followed by a complement.
On the other hand, Sentences (23) and (25) are ambiguous, containing
a main clause followed by either a complement or a relative. The
ambiguity in (23) involves the distinction, for example, of whether
John's reporting was regretted (complement) or whether the fact which
he happened to report was regretted (relative). Speakers were instructed
to read sentences (23) and (25) on their relative clause meanings.
Measurements of the duration of the key word fact were made for
these sentences using 7 speakers. The average durations for sentences
(22)-(25) were, respectively, 261.6 msec, 269.6 msec, 250.6 msec, and
273.0 msec. All 7 speakers produced a longer duration in the relative-
containing (23) than in the complement-containing (24). However,
remaining comparisons between complement- and relative-containing
sentences failed to uphold this highly systematic lengthening effect
for relatives. If the greater lengthening for relative-containing
sentences turns out to be significant on the basis of further experi-
mentation, thib effect can be accounted for by the notion that a
grammatical boundary of greater hierarchical status occurs immediately
after fact in the case of relatives.
5. Concluding Remarks
If this chapter serves its purpose, it leaves the unmistakable
156
impression of a job unfinished. A major reason for undertaking this
work on speech timing was the suspicion that the work would not, and
should not, end with the completion of this thesis. The study overall
is viewed as a programmatic series of steps, only a few of which have
been taken.
In addition to the experiments described above, there exist
at least three major issues in timing research on which a great deal
of theoretical and empirical work is required. By way of concluding
the discussion, these issues are sketched below with reference to the
findings of previous chapters.
a. Syntax vs. Rhythm. One of the issues that has cropped up
from time to time in the current work on speech timing is the problem
of determining whether a specific durational effect should be attributed
to a level of syntactic processing or to a level of phonological
processing that is concerned with rhythmic factors such as stress
timing. In some cases, it is possible to construct experiments that
adequately test the effects of a syntactic variable independently of
rhythmic variables. However, in other cases, it is difficult to
achieve this controlled situation; a change in the value of a syntactic
variable may be typically or invariably accompanied by a perceptible
change in speech rhythm. An example of a case in which the two variables
typically covary concerns the deletion Conjunction Reduction as used in
Chapter 2. In the relevant experiment, a comparison was attempted
between a two-clause sentence in which the subject of the second clause
was either present or deleted from surface structure. Deletion of the
subject (a change in a syntactic variable) was accompanied by a change
in the number of stressed syllables in the vicinity of the major clause
157
boundary, and it is conceivable that the latter change produced a
reorganization of rhythmic stress affecting the duration of the key
segment, occurring just prior to the clause boundary. In order to
determine whether syntactic or rhythmic factors (or both) were
responsible for the obtained effect, two types of test need to be
conducted. First, a test could be constructed in which the difference
in speech rhythm between the comparison sentences is minimized, in
this case, by substituting an unstressed pronoun for a stressed noun
in the subject position of the second clause in the non-deletion
sentence. Second, an independent test could be constructed in order
to test the effects of rhythmic stress pattern independently of
syntactic variables. The latter type of test is of course particularly
important in cases where the first type of test cannot br. constructed,
as may well be the case for some syntactic constructions. Work along
these lines should help in the development of separate theories of
syntactic and rhythmic phonological influences on speech timing.
b. Information-Flow. A concern related to the issue about a
possible rhythmic level of processing is the more general question of
what distinct processing stages actually influence speech timing, and
how is the flow of information among these stages organized. So far,
two distinct types of non-phonetic timing effects have been studied,
including clause-final lengthening and antecedent lengthening. As
noted earlier, it is believed that these two types of lengthening can
be distinguished empirically on the basis of their domain of control.
Clause-final lengthening, according to any of its major alternative
explanations (see Chapter 1), is considered to be a reflection of the
speaker's clause-by-clause computation (see Cooper, in preparation,
158
for detailed assumptions of this viewpoint). Antecedent lengthening,
on the other hand, can be controlled by a coreferent that belongs to
a different syntactic clause, eliminating the possibility that non-
phonetic timing control takes place exclusively at a level of computation
restricted to the scope of a single clause.3
Up to this point, we have at least two distinct levels of non-
phonetic timing control. There may be more levels, but for the present,
it is appropriate to ask whether these two levels operate in series or
in parallel. Intuitively, it is appealing to hypothesize that semantic
effects like antecedent lengthening (as well, perhaps, as effects like
emphatic and contrastive stress assignment), are programmed before the
speaker's computation is restricted to clause-by-clause scope. A
behavioral test of this hypothesis is currently difficult but not
impossible to construct. The essential ingredients for such a test
would be two separate lengthening effects, one akin to antecedent
lengthening, the other clause-final lengthening, with each producing
a large lengthening effect, say 30%. If the two 30% effects are
combined in the same sentence, and if these effects add linearly, then
a parallel model of lengthening predicts a total increase in length of
60%, whereas a serial model predicts a somewhat larger increase,
amounting to 69%. Further work on timing effects and their interaction
might allow us to conduct an experiment of this sort to test the
seriality hypothesis. In any event, it appears that further work on
timing should be directed at providing a better model of information-
flow during sentence production.
159
c. Explanations of Clause-Final Lengthening. Finally,
we return to an issue raised in Chapter 1 concerning possible explana-
tions of clause-final lengthening, the effect that has occupied most
of our experimental concern thus far. An explanation of the effect
is desirable primarily because it might provide information about the
operation of the level of timing constrained by clause-clause
computation. Three major types of explanation were offered earlier,
concerning the speaker's planning strategy, the speaker's storage
and retrieval of clauses, and a listener-oriented account, whereby
clause-final lengthening is produced in order to aid the listener
in recovering structural information. Based on the results of the
present study, it is possible to reject a least one of these major
accounts for certain constructions. Recall that the average magnitude
of lengthening was no more than 10 msec for some structural distinctions.
In these cases, a listener-oriented account can be ruled out, since
listeners are unable to reliably detect such small differences in
duration during sentence perception (Klatt and Cooper, 1975).4 We are
left then, with the planning and buffer accounts. Tentatively, it
seems that the buffer account should be favored, since it is far from
clear that a 10 msec-interval would be sufficiently large to be of any
advantage to clause planning. But only when much more is known will we be
in a position to test the merits of the planning and buffer accounts
adequately.
160
References
Borkin, A. (1973). To be and not to be. in C. Corum, T.C. Smith-Stark,
and A. Weiser (eds.) Payers from the Ninth Regional Meeting of the
Chicago Linguistic Soci Chicago: Chicago Linguistic Society.
Pp. 44-56.
Cooper, W.E. (in preparation). Syntactic Control of Speech Timing.
Jackendoff, R.S. (1975). Morphological and Semantic Regularities in the
Lexicon. Language 51, 639-671.
Postal, P.M. (1974). On Raising: One Rule of English Grammar and Its
Theoretical Implications. Cambridge: M.I.T. Press.
Vetter, D.C. (1973). Someone solves this problem tomorrow. Linguistic
Inquiry 4, 104-108.
161
Footnotes
I am grateful to John Robert Ross for providing this example.
2The comparison between complement and relative clauses was suggested
to me by Merrill Garrett.
3The results of Experiment 6 are not easily handled by the distinction
drawn here. For further work on this problem, see Cooper (in prepara-
tion).
4It is important to note that the smallness in magnitude of some of the
present effects does not detract from their potential value in aiding
in an understanding of the speech production system. It is true,
however, that such small-magnitude effects are of less interest than
large-magnitude effects from the standpoint of practical applications,
including speech synthesis.