Modeling the perception of speaker age and sex in children's voices

Abstract

At previous meetings, we presented data on the perception of speaker sex and age in children's voices. The stimuli common to these experiments were /hVd/ syllables in isolation and sentence context. Here we present the results of a modeling study in which acoustic measurements of the /hVd/ syllables were used to predict listener judgments of age and sex. Variables were selected based on preliminary analyses and suggestions from the literature: (1) duration; (2) average fundamental frequency; (3) geometric mean of F1 F2 F3; (4) magnitude difference between harmonics 1 and 2; (5) magnitude difference between harmonic 1 and F3 peak; (6) Cepstral pitch prominence; (7) Harmonic-to-noise ratio. Logistic regression models were constructed to predict listeners' judgments of speaker sex, and mixed effects linear regression models for speaker age. Results confirmed the importance of F0, formant frequencies and measures related to the voicing source for both age and sex. Regression coefficients for judgments of age and sex were similar to those for veridical age and sex when regressed on the same physical measures, suggesting a near-optimal use of cues by listeners.