Spectral subtraction for reverberation reduction applied to automatic speech recognition

Abstract

In real conditions, reverberation severely compromises the accuracy of automatic speech recognition (ASR) systems. The use and the assessment of spectral subtraction for reducing the reverberation effect in ASR systems is discussed. For such, a general procedure is presented. The influence of some important dereverberation parameters, such as the reverberation time, on the recognition rate is evaluated. By using the referred approach, the speech recognition rate is considerably improved.