Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Speech Recognition in Noisy Reverberant Rooms Using a Frequency Domain Blind Deconvolution Method

G. Nokas, E. Dermatas

Wire Communications Laboratory, Electrical & Computer Engineering Department, University of Patras, Greece

The aim of this paper is to present and evaluate two adaptive speech enhancement methods in the frequency domain by measuring the recognition rate of a speaker-independent word recognition system of isolated words. In a hands-free speech recognition experiment, a factory noise source and a speaker are the acoustic sources in a real room environment. A close-talking microphone is positioned near the noise source while the primary omni-directional microphone captures the convoluted speech and noise signals. Adaptive noise cancellation reduces the presence of noise in the primary microphone followed by a blind deconvolution algorithm used to minimize reverberations. Experimental results showed that the proposed speech enhancement methods increase 2.4 times the recognition rate for a vocabulary of 21 words of the Greek language, but the achieved recognition rate of 16% is inapplicable for commercial applications.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Nokas, G. / Dermatas, E. (1999): "Speech recognition in noisy reverberant rooms using a frequency domain blind deconvolution method", In EUROSPEECH'99, 2853-2856.