Inter-frequency dependency in mmse speech enhancement
Chunjian Li, S.V. Andersen · 2004
In this paper an MMSE estimator of the complex shortime spectrum is considered for optimum noise reduction of speech. The correlation between frequency components is exploited to improve the estimation, especially of those components with low local SNR. Furthermore, by making use of both spectral envelope and time envelope, the estimator is able to suppress noise power infrequency domain and lime domain simultaneously. The performance of the resulting estimator is found to be superior to the non-causal IN Wiener filter The enhanced signal suffers less spectral distortion, while achieving a lower mean squared error than the Wiener filter.