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

In this paper, the problem of deteriorating performance of speech recognition under very low signal-to-noise ratios (SNR) is considered. In particular, for a given pre-trained speech recognizer and for a finite set of speech commands, we show that popular noise reduction methods have a mixed performance in speech recognition accuracy under very low SNR. Although most noise reduction methods are attempting to reduce speech distortion or to increase noise suppression, it does not necessarily improve speech recognition accuracy very much due to the complexity of the recognizer. We propose a new hybrid algorithm to optimize on the speech recognition accuracy directly by mixing different noise reduction methods together. We show that this method can indeed improve the accuracy significantly.