Speaker Adaptation Based on a Maximum Observation Probability Criterion

概要

論文の詳細を見る
A speaker adaptation technique that maximizes the observation probability of an input speech is proposed. It is applied to semi-continuous hidden Markov model (SCHMM) speech recognizers. The proposed algorithm adapts the mean μ and the covariance Σ iteratively by the gradient search technique so that the features of the adaptation speech data could achieve maximum observation probabilities. The mixture coefficients and the state transition probabilities are adapted by the model interpolation scheme. The main advantage of this scheme is that the means and the variances, which are common to all states in SCHMM, are adapted independently from the other parameters of SCHMM. It allows fast and precise adaptation especially when there is a large acoustic mismatch between the reference model and a new speaker. Also, it is possible that this scheme could be adopted to other areas which use codebook. The proposed adaptation algorithm was evaluated by a male speaker-dependent, a female speaker-dependent, and a speaker-independent recognizers. The experimental results on the isolated word recognition showed that the proposed adaptation algorithm achieved 46.03% average enhancement in the male speaker-dependent recognizer, 52.18% in the female speaker-dependent recognizer, and 9.84% in the speaker-independent recognizer.
社団法人電子情報通信学会の論文
2001-02-01