Masked noise probability-based speech enhancement

An optimal approach for enhancing speech signals degraded by uncorrelated stationary additive noise with auditory perception properties is proposed. Based on auditory masking effects, the speech spectra estimates include noisy speech spectra of masked noise and a classical spectral subtraction estimate of unmasked noise. Taking into account the uncertainty of the noise presence, the enhanced speech signal spectra are calculated by a weighted sum of these two estimates and the weights are derived from the masked noise probability (MNP). The performance of the proposed speech enhancement approach has been evaluated with speech distortion and informal listening tests. Comparison with Azirani´s method and the short-time spectral amplitude estimate based on minimum mean square error (MMSE-STSA) shows that the speech distortion apparently decreased and the musical noise was suppressed.