Maximum likelihood linear regression (MLLR) speaker adaptation using dynamic weighting

摘要

According to the prior art Maximum Likelihood Linear Regression strategies have in common that the influence or weight of a new utterance remains the same throughout the whole adaptation process. According to the present invention, after a first very quick adaptation to a new speaker, new utterances are weighted less than all previous speaker specific utterances so that the sum of older utterances from a specific speaker have much more influence than a few new ones, if this speaker uses the system for a long time.