Perceptually Motivated Generalized Spectral Subtraction for Speech Enhancement

This paper addresses the problem of single speech enhancement in adverse environment. The common noise reduction techniques are limited by a tradeoff between an efficient noise reduction, a minimum of speech distortion and musical noise.in this work, we propose a new speech enhancement approach based on non-uniform multi-band analysis. The noisy signal is divided into a number of sub-bands using a gammachirp filter bank with non-linear ERB resolution, and the sub-bands signals are individually weighted according the generalized spectral subtraction technique. For evaluating the performance of the proposed speech enhancement, we use the perceptual evaluation measure of speech quality PESQ and the subjective quality rating designed to evaluate speech quality along three dimensions: signal distortion, noise distortion and overall quality. Subjective evaluation tests demonstrate significant improvements results over classical subtractive type algorithms, when tested with speech signal corrupted by various noises at different signal to noise ratios.