A multi-filter system for speech enhancement under lowsignal-to-noise ratios

In this paper, the problem of deteriorating performance of speechrecognition under very low signal-to-noise ratios (SNR) isconsidered. In particular, for a given pre-trained speech recognizerand for a finite set of speech commands, we show that popular noisereduction methods have a mixed performance in speech recognitionaccuracy under very low SNR. Although most noise reduction methodsare attempting to reduce speech distortion or to increase noisesuppression, it does not necessarily improve speech recognitionaccuracy very much due to the complexity of the recognizer. Wepropose a new hybrid algorithm to optimize on the speech recognitionaccuracy directly by mixing different noise reduction methodstogether. We show that this method can indeed improve the accuracysignificantly.