An all-phoneme ergodic HMM for unsupervised speaker adaptation

The author proposes an all-phoneme ergodic HMM (hidden Markov model) that incorporates stochastic language constraints in unsupervised speaker adaptation. The proposed model consists of all-phoneme HMMs and interphoneme probabilities. It can be regarded as a rather large single ergodic HMM containing hidden states of all phonemes as well as intraphoneme and interphoneme transition probabilities. Since this model is a model of arbitrarily spoken words, the standard Baum-Welch reestimation algorithm can be used to train the whole ergodic model. In the experiments, only mean vectors of the state output probability densities are reestimated, and a vector field smoothing algorithm is used to enhance the statistical reliability. The proposed method was tested on phoneme and phrase recognition experiments with male reference and input speakers. A better performance than with the speaker-independent case was attained by using adaptation data shorter than three minutes.<>