Fully adaptive SVD-based noise removal for robust speech recognition

This paper presents a new approach to improve the robustness of large vocabulary continuous speech recognition. The proposed technique { based on Singular Value Decomposition (SVD) { originates from classical signal enhancement, but it is adapted to the speci c requirements imposed by the speech recognition process. Additive noise reduction is obtained by altering the singular value spectrum of the signal observation matrix, thereby preserving speech signal components and suppressing noise-related components. The basic algorithms are developed for white noise but they can easily be extended to the general coloured noise case. With the aid of a noise reference, non-stationary noise can be handled as well. All schemes are adaptive, and work in real-time. Recognition experiments on a noise-corrupted database with large vocabulary, continuous speech (Resource Management) reveal that relative reductions of the WER of more than 60% are obtained.