ITRW on Adaptation Methods for Speech Recognition August 29-30, 2001 Sophia Antipolis, France

Jacobian Approach to Joint Adaptation to Noise, Channel and Vocal Tract Length

Shigeki Sagayama (1,2), Yutaka Kato (2), Mitsuru Nakai (2), and Hiroshi Shimodaira (2)

(1) The University of Tokyo, Bunkyo-ku, Tokyo, Japan
(2) Japan Advanced Institute of Science and Technology, Ishikawa, Japan

This paper describes the Jacobian approach to simultaneously adapting acoustic models to unknown noise, channel and vocal tract length from a supervised adaptation data. As has been both theoretically and experimentally shown, Jacobian adaptation is one of most efficient methods for model adaptation if the target condition is close to the initial condition. It utilizes the linear relationship in the neighbor of the initial condition which in turn can be used in decomposition of multiple factors. The analytic relationship between noise, channel, vocal tract length and the observed cepstrum is linearized using the Jacobian matrices. Least squares fit gives the estimates of noise, channel and vocal tract stretch parameters. Experimental evaluation gave a significant improvement to the recognition accuracy.

Full Paper

Bibliographic reference.  Sagayama, Shigeki / Kato, Yutaka / Nakai, Mitsuru / Shimodaira, Hiroshi (2001): "Jacobian approach to joint adaptation to noise, channel and vocal tract length", In Adaptation-2001, 117-120.