Speech Recognition by Denoising and Dereverberation Based on Spectral Subtraction in a Real Noisy Reverberant Environment

A blind dereverberation method based on spectral subtraction using a multi-channel least mean squares algorithm was previously proposed. The results of a large vocabulary continuous speech recognition task showed that this method achieved significant improvements over the conventional method based on cepstral mean normalization and beamforming in a simulated reverberant environment without additive noise. In this paper, we evaluate the blind dereverberation method in a real noisy reverberant environment. We present a denoising and dereverberation method based on power spectral subtraction or generalized spectral subtraction, and evaluate our proposed method using speech in a real environment. The generalized spectral subtraction based method achieves an average relative word error reduction rate of 39.1% and 11.5% compared to the conventional cepstral mean normalization and power spectral subtraction based methods, respectively.