Research on speech enhancement based on MMSE-LSA

In this paper, the speech enhancement algorithm based on minimum mean square error short time log spectral amplitude estimation (MMSE-LSA) is studied. The estimates of prior SNR and noise are the key points of MMSE-LSA. As to the estimate of noise, we use the smoothing parameters to update the noise estimate in speech-absent and speech-present frame respectively and this method can well adapt to the non-stationary environment. In order to suppress the residual noise effectively while reduce the chance of distortion of speech, noise masked probability is calculated and combined with the speech enhancement algorithm. MMSE-LSA has shown to be asymptotically near optimal for speech signals corrupted by additive white noise. However, for some narrow band noise, the noisy speech signals need to be whitened to improve the speech enhancement performance.