Variability in voice spaces for individual talkers

Our previous studies of psychoacoustic variability in voice [e.g., J. Acoust. Soc. Am. 151, 3462–3472, 2022] have shown that two principal components, corresponding approximately to arousal and perceived physical size, account for 30–35% of psychoacoustic variability within and across talkers. Beyond these, principal component analyses (PCA) consistently identify five to six additional factors, which appear to differ across individuals and may reflect what makes each voice unique. In this study we examined these higher-order components in detail to determine whether their variation is idiosyncratic or whether shared structure and patterns exist. Preliminary results suggest that while these later components do weigh on different acoustic variables for different talkers, the patterns of weighting are not random. Instead, small subsets of talkers appear to share one of a limited number of recurring configurations for each PC, suggesting that individual voice identities may draw from a constrained repertoire of multidimensional acoustic profiles. These findings indicate that the relationship between psychoacoustic voice spaces for different talkers is both orderly and highly complex. We discuss implications of these complexities for models of vocal similarity, speaker identification, and voice recognition by both humans and machines.