Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Fast Speaker Adaptation Through Spectral Transformation for Continuous Speech Recognition

H. C. Choi, R. W. King

Speech Technology Research Group, Department of Electrical Engineering, University of Sydney, Australia

In this paper, the use of spectral transformation to perform speaker adaptation for continuous speech recognition is presented. Speaker adaptation is performed by transforming the feature vectors of the speech of a new speaker to the spectral space of a reference speaker who has participated in the training of a reference recognition system. The same reference recognition system is used for all new speakers but the individual transformation for each new speaker is estimated accordingly by canonical correlation analysis (CCA). Since this approach does not require any modifications to the reference system, itis especially attractive for the development of on-line adaptive recognition systems which require fast adaptation. The effectiveness of this approach is investigated by performing adaptation experiments using the DARPA 1000-word Resource Management (RM1) continuous speech corpus.

Full Paper

Bibliographic reference.  Choi, H. C. / King, R. W. (1994): "Fast speaker adaptation through spectral transformation for continuous speech recognition", In ICSLP-1994, 2219-2222.