Robust speech recognition using singular value decomposition based speech enhancement

Abstract

Speech recognition systems work reasonably well in laboratory conditions, but their performance deteriorates drastically when they are deployed in practical situations where the speech is corrupted by additive noise. One way to improve the performance of a speech recognition system in the presence of noise, is to enhance the speech prior to its recognition. Two singular value decomposition based techniques have been recently proposed for speech enhancement [5] [6]. In these techniques, singular value decomposition has been applied to an over-determined, over-extended data matrix formed from the noisy speech signal. A noise-free, low rank approximation was obtained by retaining a specific number of singular values. This technique was applied here as a preprocessor for recognising speech in the presence of noise. It was found to improve the recognition performance significantly for signal-to-noise ratios less than 15dB.

Speech enhancement using the multistage Wiener filter

“…The MWF [1] approximates the Wiener filter from a Krylov subspace. Contrary to [6], [7], the MWF does not require knowledge of the full covariance matrices, and it does not employ eigenvalue or singular value decomposition. The MWF tends to achieve nearly the Wiener performance using a reduced rank filter and has been successfully applied to numerous problems, specifically in Radar Detection [8]- [10] and communications [11]- [13].…”

“…The Wiener filter coupled with subspace reduction (see, e.g., [6], [7]), have performed surprisingly well in this problem. The MWF [1] approximates the Wiener filter from a Krylov subspace.…”