A new gain model for speech enhancement

A new gain model for estimating speech is presented in this paper, in which there are the same parameters as MMSE-LSA gain model, but there is a different form. This new gain model prevails over the model based on maximum likelihood estimation in restraining musical noise, and is superior to MMSE-LSA gain model in both preserving weak speech components and tracking the starts and ends of speech. Further more, by introducing the future frame into averaging the original power spectrum and by modifying the undervalued noise, the algorithm based on the new gain model gets a more satisfactory performance. Experimental results demonstrate the feasibility and effectivity of the algorithm.