Automatic classification of consonant-vowel transitions based on subglottal resonances and second formant frequencies.
Nancy F. Chen, Steven M. Lulich · The Journal of the Acoustical Society of America · 2009
A growing body of work indicates that subglottal resonances (SGRs) divide the frequency space of vowels and consonants into distinct regions in several languages [S. Lulich, J. Phonetics (in press), and references therein]. For instance, the second formant frequency (F2) of back vowels lies between the first and second subglottal resonances (Sg1 and Sg2), while for front vowels F2 lies at higher frequencies than Sg2. In alveolar stop bursts, F2 lies between Sg2 and the third subglottal resonance (Sg3). In this study, we build on previous work [S. Lulich, JASA 124, 2558 (2008)] to automatically determine SGRs using locus equations computed from 405 consonant-vowel (CV) transitions produced by an adult male native speaker of American English. Consonants and vowels are then classified into place of articulation and front/back features based only on the relations between F2 and the SGRs. The mean hit rate across categories was 77%, while the mean false alarm rate was 12%. These results show relative improvements of miss and false alarm rates by 21% and 14% from a baseline experiment, where F2 without SGRs was used to perform the classification. Our knowledge-based approach of classifying CV transitions is potentially useful when vowel and consonant identities are unavailable.