Building talker knowledge from talker variability

Abstract

Listeners reliably extract invariant linguistic category information from speech despite massive cross-talker variability. Existing work shows that talker variability is in part handled by the learning and maintaining of category-specific acoustic distributions towards the statistics in the spoken input from a talker. When do listeners maintain information about talkers’ speech after we meet them, and what kind of information should be maintained in principle? In this talk, I present a computational framework that addresses these questions by linking the statistics of the speech input (production) to predictions about perception. The results provide constraints on the type of inference underlying talker-related adaptivity during speech perception and have direct implications for current research on talker recognition.