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

A Generalized Dynamic Cepstrum for Hands-Free Speech Recognition

Hiroshi Matsumoto, Yoshihiro Ito, Akihiko Shimizu and Kazumasa Yamamoto

Faculty of Engineering, Shinshu University, Wakasato, Nagano, Japan

This paper examines the effectiveness of a generalized dynamic cepstrum in hands-free speech recognition. The generalized dynamic cepstrum (DyMFGC) is based upon the forward masking on the generalized logarithmic spectrum instead of the log-spectrum, which intends to make robust to additive noise as well as convolutional noise. Digit recognition tests are carried out under a relatively quiet and small size office environment. Under white noise environments, the DyMFGC outperforms the dynamic cepstrum on the logarithmic spectrum and MFCC with cepstral mean normalization, and maintains the word accuracy of 90% to 95% within a 1 m distance from a source. In speech babble noise environments, the performance of the DyMFGC is approximately same as that of the dynamic cepstrum on the logarithmic amplitude scale.

Full Paper

Bibliographic reference.  Matsumoto, Hiroshi / Ito, Yoshihiro / Shimizu, Akihiko / Yamamoto, Kazumasa (2001): "A generalized dynamic cepstrum for hands-free speech recognition", In HSC2001, 115-118.