Reliable cues to gender and talker identity are present in a short vowel segment recorded in running speech

Abstract

Based on the source-filter model, individual variation in vocal-tract anatomy has long been proposed to influence the acoustic realization of phonetic gestures. It is therefore noteworthy that humans are sexually dimorphic, with individuals being uniquely marked through variation in many anatomical features. Both observations apply to vocal-production structures, and sounds whose characteristics produce a high-fidelity ‘‘off-print’’ of underlying anatomy should therefore provide clear cues to gender and identity. Vowels are arguably optimal in this regard, as they reflect details of both source and filter components. This conception was tested using 15 instances of the utterance ‘‘Test n Test’’ recorded from each of 50 males and 75 females. Measurements of the vowel in the first ‘‘Test’’ were entered in discriminant-function classification. Using only fundamental frequency and a formant-based estimate of vocal-tract length, 1852 of 1875 (99%) tokens were correctly classified by talker gender. Using all variables, 593 of 750 (79%) tokens recorded from males were sorted by talker, while the comparable figures were 729 of 1125 (65%) for females. These results underscore the importance of differences in vocal- tract anatomy for individual variation in speech production and suggest a prominent role of vocalic sounds in indexical cuing. [Work supported by NIMH.]