Speaker-independent vocal tract shapes that optimally elicit vowel recognition

When speakers shape their vocal tracts in specific ways, it elicits the recognition of certain phonemes in listeners. But, which vocal tract shapes, independent of any particular speaker, best explain the phoneme recognition behavior of listeners? Prior studies have associated recognition with simplified shaping actions—e.g., isolated, narrow constrictions. In this study, we use the method of Reverse Correlation (RC) to reconstruct unconstrained vocal tract area functions that best explain the phoneme recognition behavior. Here, we ask listeners to render yes–no recognition responses over a series of ambiguous stimuli derived from random area functions, which are then used to estimate an area function via RC regression analysis to optimally predict the listener responses. Using RC, optimal area functions were reconstructed for vowels /i/, /a/, and /u/ from native English speakers with normal hearing (4 collected, 30 planned). Current results suggest that these area functions accurately predicted listener responses (67%–90%, mean 78%). Furthermore, reconstructed area functions reveal specific localized constrictions of the vocal tract that contribute consistently to explaining responses across listeners, but which were different from constrictions exhibited by speakers producing those same vowels. Results hold implications for understanding the complex relationship between perception and production of speech.