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

Novel Noise Rejection Using Speechness Score Based on Frame-Level Confidence Measurement

S. Taniguchi, Y. Obuchi, A. Akio, and N. Hataoka

Central Research Laboratory, Hitachi Ltd., Kokubunji, Tokyo, Japan

We define "speechness" as the possibility for speech sound in this paper. To separate speech and noise, or non-speech, is one of the important issues for speech recognition applications. We developed a noise rejection algorithm based on novel confidence measurement for speechness. This rejection algorithm is compact enough to implement into embedded systems [1].

We performed the evaluation experiments on 2.5-hour noise data and 3200-isolated-word database. As the evaluation results, a noise rejection ratio was 78%, a recognition rate decreased by only 0.14%. The rejection algorithm needs 2 kbytes memory additionally in our system, and the response time of the speech recognition increased by only 4.4%.

Full Paper

Bibliographic reference.  Taniguchi, S. / Obuchi, Y. / Akio, A. / Hataoka, N. (2001): "Novel noise rejection using speechness score based on frame-level confidence measurement", In HSC2001, 179-182.