Subword-based approaches for spoken document retrieval

Abstract

This thesis explores approaches to the problem of spoken document retrieval (SDR), which is the task of automatically indexing and then retrieving relevant items from a large collection of recorded speech messages in response to a user specified natural language text query. We investigate the use of subword unit representations for SDR as an alternative to words generated by either keyword spotting or continuous speech recognition. Our investigation is motivated by the observation that word-based retrieval approaches face the problem of either having to know the keywords to search for [\em a priori], or requiring a very large recognition vocabulary in order to cover the contents of growing and diverse message collections. The use of subword units in the recognizer constrains the size of the vocabulary needed to cover the language; and the use of subword units as indexing terms allows for the detection of new user-specified query terms during retrieval. Four research issues are addressed. First, what are suitable subword units and how well can they perform? Second, how can these units be reliably extracted from the speech signal? Third, what is the behavior of the subword units when there are speech recognition errors and how well do they perform? And fourth, how can the indexing and retrieval methods be modified to take into account the fact that the speech recognition output will be errorful?

(cont.) We first explore a range of subword units ofvarying complexity derived from error-free phonetic transcriptions and measure their ability to effectively index and retrieve speech messages. We find that many subword units capture enough information to perform effective retrieval and that it is possible to achieve performance comparable to that of text-based word units. Next, we develop a phonetic speech recognizer and process the spoken document collection to generate phonetic transcriptions. We then measure the ability of subword units derived from these transcriptions to perform spoken document retrieval and examine the effects of recognition errors on retrieval performance. Retrieval performance degrades for all subword units (to 60% of the clean reference), but remains reasonable for some subword units even without the use of any error compensation techniques. We then investigate a number of robust methods that take into account the characteristics of the recognition errors and try to compensate for them in an effort to improve spoken document retrieval performance when there are speech recognition errors. We study the methods individually and explore the effects of combining them. Using these robust methods improves retrieval performance by 23%. We also propose a novel approach to SDR where the speech recognition and information retrieval components are more tightly integrated.

(cont.) This is accomplished by developing new recognizer and retrieval models where the interface between the two components is better matched and the goals of the two components are consistent with each other and with the overall goal of the combined system. Using this new integrated approach improves retrieval performance by 28%. ...