Accent adaptation of the recognition model.

A) The cochleagrams represent two utterances of “eight”. A module originally learned the word “eight” spoken with a British (North England) accent (top) and then recognized an “eight” spoken with a New Zealand accent (bottom). B) The module trained on the British accent was allowed to adapt to the New Zealand accent with the corresponding precision values for the first level sensory (causal) and internal (hidden) states (sensory log-precision: and internal log-precision: where from left to right). For each precision ratio, we plotted the reduction in prediction error (of the causal states, see Model) after five repetitions of the word “eight” spoken with a New Zealand accent. As expected, accent adaptation was accomplished only with high sensory/internal precision ratios (resulting in greatly reduced prediction errors) whereas no adaptation occurred (prediction errors remained high) when this ratio was low.