First International Conference on Spoken Language Processing (ICSLP 90)
A novel speaker adaptation method, which is applied to the demi-syllable based speech recognition system using continuous density HMM, is proposed. In this method, mean vectors of HMM Gaussian pdfs for a standard speaker are adapted to those for a new speaker with a small amount of training data. Supervised speaker adaptation is first employed, and for the recognition units which are not adapted in the supervised adaptation, unsupervised speaker adaptation is performed. The effectiveness of the proposed method was confirmed by large vocabulary word recognition experiments. Using 50 word utterances for speaker adaptation, the recognition rates were improved by 14.4 %, on an average.
Bibliographic reference. Shinoda, Koichi / Iso, Ken-ichi / Watanabe, Takao (1990): "Speaker adaptation for demi-syllable based speech recognition using continuous HMM", In ICSLP-1990, 261-264.