Vocal Tract Disparity and Potential Implications for Speaker Recognition

Abstract

Perceptual experiments show that listeners recognize female speakers with lower accuracy than male speakers. Automatic speaker recognition systems may also show performance bias against female speakers even when training data sets are gender balanced. The underlying reasons for this discrepancy are unclear. Here, we apply geometric morphometrics to quantify sex-related morphological vocal tract disparity -- the extent of shape variation -- across both resting and articulatory configurations. We find that male speakers exhibit greater disparity in both resting and articulatory configurations. This morphological idiosyncrasy may in turn generate more discriminable acoustic signatures and offer a biological explanation for higher recognition accuracies for male voices by humans and machines. Our results suggest that innate variation in vocal tract morphology may contribute to performance bias in voice technology and voice perception by human listeners.