Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Fully Adaptive SVD-based Noise Removal for Robust Speech Recognition

Kris Hermus, Ioannis Dologlou, Patrick Wambacq, Dirk Van Compernolle

Katholieke Universiteit Leuven - ESATPSI, Belgium, and Lernout & Hauspie Speech Products, Belgium

This paper presents a new approach to improve the ro-bustness of large vocabulary continuous speech recogni-tion. The proposed technique { based on Singular ValueDecomposition (SVD) { originates from classical signalenhancement, but it is adapted to the specific require-ments imposed by the speech recognition process.Additive noise reduction is obtained by altering the sin-gular value spectrum of the signal observation matrix,thereby preserving speech signal components and sup-pressing noise-related components.The basic algorithms are developed for white noise butthey can easily be extended to the general coloured noisecase. With the aid of a noise reference, non-stationarynoise can be handled as well. All schemes are adaptive,and work in real-time.Recognition experiments on a noise-corrupted databasewith large vocabulary, continuous speech (Resource Man-agement) reveal that relative reductions of the WER1ofmore than 60 % are obtained.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Hermus, Kris / Dologlou, Ioannis / Wambacq, Patrick / Compernolle, Dirk Van (1999): "Fully adaptive SVD-based noise removal for robust speech recognition", In EUROSPEECH'99, 1951-1954.