Speech enhancement using both spectral and spectral modulation domains

This paper proposes a speech enhancement approach that uses both the spectral and spectral modulation domains. In this approach, the noisy speech signal is enhanced simultaneously in the spectral domain, using a minimum mean square error (MMSE) short time spectral amplitude estimator, and in the spectral modulation domain, using a MMSE spectral modulation magnitude estimator. The results of both estimators are then weighted and combined together, using a function based on the a posteriori SNR, to produce the desired enhanced signal. Comparative results using both the segmental SNR and PESQ objective measures are presented for both stationary and non-stationary noises. It is observed that the proposed approach suppresses more noise than the compared approaches, but at the usual compromise of introducing speech distortions.