International Workshop on Hands-Free Speech Communication (HSC2001) April 9-11, 2001 Kyoto, Japan

Environment-adaptive algorithms for robust speech recognition

Jean-Claude Junqua (1), C. Cerisara (2), L. Rigazio (1), and D. Kryze (1)

(1) Panasonic Speech Technology Laboratory, Santa Barbara, CA, USA
(2) LORIA UMR, Vandoeuvre-Ies-Nancy, France

After reviewing some desired characteristics of environment-adaptive algorithms, we describe a joint noise and channel adaptation algorithm that has the merit of being computationally efficient. This algorithm, based on a first-order approximation, takes its roots in the Vector Taylor Series (VTS) and Jacobian adaptation methods. Digit recognition experiments for data recorded in a car and simulated channel distortions show that this algorithm is very effective in dealing with both types of mismatch. By applying dimensionality reduction techniques such as Principal Component Analysis (PCA) it is possible to reduce significantly its computational complexity without any decrease in recognition accuracy. When this algorithm is used in combination with a robust front-end based on subband analysis we show tbat, for best performance, the front-end and the environment-adaptive algorithm have to be jointly optimized.

Full Paper

Bibliographic reference.  Junqua, Jean-Claude / Cerisara, C. / Rigazio, L. / Kryze, D. (2001): "Environment-adaptive algorithms for robust speech recognition", In HSC2001, 31-34.