Speaker Adaptation Based on MAP Estimation Using Fuzzy Controller

Abstract

This paper presents a technical incremental speaker adaptation method called FCMAP, which incorporated an appropriate fuzzy controller into maximum a posteriori (MAP) to adapt the hidden Markov model parameters. The recognition performance of MAP is consistently improved and stabilized by an appropriate fuzzy controller. The recognition results obtained in unsupervised adaptation experiments showed that FCMAP estimation was effective even when only few utterances from a new speaker were used for adaptation.