First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Maximum Likelihood Estimation of Speech Waveform Under Nonstationary Noise Environments

Kaoru Gyoutoku, Hidefumi Kobatake

Faculty of Technology, Tokyo University of Agriculture & Technology, Tokyo, Japan

This paper describes a study on noisy speech processing. The background noise is assumed to be a sum of a stationary noise and isolated nonstationary noises, and it is also assumed that there are no overlapping between speech segments and nonstationary noises. The noise processing system described in this paper consists of three subsystems. The first subsystem is to detect nonstationary segments buried in the stationary noise. The second one is for the speech/nonspeech discrimination of the detected nonstationary segments. If they are identified as speech, the third subsystem begins to work for speech enhancement. This paper gives a detailed discussion on the third subsystem, which is based on the maximum likelihood estimation method. Experimental results show that the improvement in signal-to-noise ratio by the proposed method is more than 13dB.

Full Paper

Bibliographic reference.  Gyoutoku, Kaoru / Kobatake, Hidefumi (1990): "Maximum likelihood estimation of speech waveform under nonstationary noise environments", In ICSLP-1990, 1149-1152.