Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


Spectral Subtraction for Front-End Noise Reduction in a Speech Recognizer

Carlos J. Teixeira, Isabel M. Trancoso

INESC / IST, INESC, Lisbon, Portugal representing the SUNSTAR consortium

The goal of this work is the design of a front-end processor to be used with state-of-the-art word recognizers, capable of reducing stationary noise from a speech signal transmitted over the public telephone network, to be implemented on a single DSP32C board, integrating an echo canceller module as well. The environment in which the speech signal is produced can vary from the typical office environment (with noise emerging from keyboards, air conditioning, etc. ) to public buildings, public phone booths and even homes. Besides this environmental noise, the speech signal is also corrupted by telephone line noise and its original amplitude and phase characteristics are also subject to channel distortion. The selected noise reduction method is the classic spectral subtraction algorithm. This selection was primarily motivated by its relative low complexity, in view of the real-time operation requirements. The tests performed so far with real environmental and telephone line noise signals confirm the effectiveness of integrating this module as a preprocessor for speech recognition in a public telephone network environment. Keywords: noise reduction, spectral subtraction, speech recognition.

Full Paper

Bibliographic reference.  Teixeira, Carlos J. / Trancoso, Isabel M. (1991): "Spectral subtraction for front-end noise reduction in a speech recognizer", In EUROSPEECH-1991, 499-502.