Figure 17.

Talker-specific speech perception can be formalized as inference under uncertainty conditional on talker identity. The probability that a particular cue value x (here, frication frequency) was intended to be category c (/s/ or /ʃ/) depends on the talker t that produced it, written p(c | x, t) (left). This probability is related via Bayes Rule to the talker-specific likelihood, the distribution of cues produced by talker t for each category, p(x | c, t) (middle). These distributions can be described by the parameters of the generative model, θ, such as the talker’s mean frication frequency for /s/ and /ʃ/ as plotted (right). Although we only plot two parameters, many more are required to even approximate the full generative model. Each talker can be thought of as a point in this (very high dimensional) parameter space.