Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Tree-Structured Speaker Clustering for Speaker-Independent Continuous Speech Recognition

Tetsuo Kosaka (1), Shoichi Matsunaga (1), Shigeki Sagayama (1)

(1) ATR Interpreting Telecommunications Research Labs., Kyoto, Japan (2) NTT Human Interface Labs., Kanagawa, Japan

We have already proposed a tree-structured speaker clustering method and its application to supervised speaker adaptation. This adaptation method is based on the selection of a speaker cluster from among multiple reference speaker clusters. Since the adaptation method employs cluster selection rather than parameter training, it can adapt quickly using only a small amount of training data. In this paper, we extend this method for application to unsupervised speaker adaptation and speaker-independent speech recognition. The results show that the adaptation method using short calibration speech (less than 5 sec) outperforms a speaker-independent recognition system.

Full Paper

Bibliographic reference.  Kosaka, Tetsuo / Matsunaga, Shoichi / Sagayama, Shigeki (1994): "Tree-structured speaker clustering for speaker-independent continuous speech recognition", In ICSLP-1994, 1375-1378.