5th International Conference on Spoken Language Processing
In this paper, we propose a Predictive Speaker Adaptation technique (PSA) in which speaker dependent HMM(SD-HMM) for a new speaker is predicted using adaptation utterances and a speaker independent HMM(SI-HMM). The method requires a prior training in order to estimate the parameters of the prediction function. For this purpose, we first prepare many speaker's fully trained SD-HMMs and their adaptation utterances(same for all speakers). In addition many speaker-specific BW-HMMs are built from the SI-HMM by means of Baum-Welch re-estimation on the adaptation utterances. The model-pair SD-HMM and BW-HMM for each speaker is used as training example for the input and output of the prediction function to find the speaker-independent prediction parameters. During adaptation, estimation of the new speaker's SD-HMM is carried out from his BW-HMM with the predetermined parameters. 60,000-word recognition experiments reported a word error-rate reduction of 16% when only 10 adaptation words were used.
Bibliographic reference. Tran, Dieu / Iso, Ken-ichi (1998): "Predictive speaker adaptation and its prior training", In ICSLP-1998, paper 1089.