Speech Enhancement Using Perceptual Wavelet Thresholding with the Ephraim and Malah Noise Suppressor and Auditory Masking

This paper presents a novel algorithm for single channel speech enhancement. The algorithm is based on the Ephraim and Malah noise suppressor with modified a priori signal-to-noise ratio estimator. The a priori SNR is biased depending on the masking threshold used for that particular frequency bin. The algorithm is performed in the wavelet domain instead of the Fourier domain. We match the sub-band of the decomposition to the critical band used in perceptual auditory modeling. Our simulation shows that the proposed algorithm is able to reduce in-band noise significantly while minimally distorting the speech signal