Romain Brunias, Jun-ichi Takami
Speaker Adaptation using
Context-dependent Continuous Density
Hidden Markov Models
Abstract:This report proposes a method of speaker adaptation using Pec-based
environment-dependent continuous density single gaussian Hidden Markov
Models. This method of adaptation is based on a state-by-state modification of the mean vectors of a set of environment-dependent HMMs, using
information extracted from adapting data (a set of words uttered by an
unknown speaker). The experiment has been implemented for a set of
environment-dependent HMMs trained with a male speaker, and this set
has been adapted by three different speakers (two males, one female).
Significant improvements of phoneme recognition rates for the unknown
speakers have been observed after using this method of adaptation.