Does perceptual learning for segmental phonotactics generalize across talkers?

In this study we present a set of experiments testing the conditions under which listeners generalize perceptual learning for biphone phonotactics across talkers. Listeners were exposed to skewed CV biphone distributions at the endpoints of a phonetic VOT continuum with more ambiguous steps having an unbiased (flat) distribution. We first replicated previous findings that listeners generalized learning from biasing contexts to perceive the higher probability biphone at ambiguous continuum steps, when all stimuli were produced by the same talker. Next, we showed that listeners also generalize learning from biphone biasing contexts learned from one talker to a new talker. Notably, there was no further improvement when listeners were presented with multiple talkers producing skewed biphone distributions. Additionally, when exposed to two talkers producing two distinct biphone distributions, listeners failed to learn talker-specific biphone probabilities, and instead evidenced learning for biphone statistics pooled across talkers. That is, listeners extracted talker-independent biphone information. We also found evidence for a possible role for talker similarity: listeners generalized somewhat more strongly between similar talkers, and when acoustic similarity for stimuli within a particular talker is greater. We argue that the results support talkerindependent perceptual learning for phonotactics, which is likely learned over representations that are rich enough to encode talker information, separately.