Automated nasal feature detection for the lexical access from features project

Abstract

The focus of this thesis was the design, implementation, and evaluation of a set of automated algorithms to detect nasal consonants from the speech waveform in a distinctive feature-based speech recognition system. The study used a VCV database of over 450 utterances recorded from three speakers, two male and one female. The first stage of processing for each speech waveform included automated 'pivot' estimation using the Consonant Landmark Detector - these 'pivots' were considered possible sonorant closures and releases in further analyses. Estimated pivots were analyzed acoustically for the nasal murmur and vowel-nasal boundary characteristics. For nasal murmur, the analyzed cues included observing the presence of a low frequency resonance in the short-time spectra, stability in the signal energy, and characteristic spectral tilt. The acoustic cues for the nasal boundary measured the change in the energy of the first harmonic and the net energy change of the 0-350Hz and 350-1000Hz frequency bands around the pivot time. The results of the acoustic analyses were translated into a simple set of general acoustic criteria that detected 98% of true nasal pivots. The high detection rate was partially offset by a relatively large number of false positives - 16% of all non-nasal pivots were also detected as showing characteristics of the nasal murmur and nasal boundary. The advantage of the presented algorithms is in their consistency and accuracy across users and contexts, and unlimited applicability to spontaneous speech.