Blind Speech Dereverberation Based on a Statistical Model

We present an algorithm to dereverberate single-channel audio signals in both noisy and noise-free acoustical environments. Recently, the model-based dereverberation that use the statistical model for room impulse responses (RIRs) is considered to be a fairly attractive approach for reverberant speech, since existing model-based estimators show the late reverberant spectral variance (LRSV) is linear-related to the previous spectral variance of noise-free reverberant signals. However, these estimators may accumulate the estimation errors of LRSV from frame to frame, especially in noisy acoustical environments. In order to improve the accuracy of the LRSV estimation, we first extend LRSV estimator to estimate the LRSV in noisy environments. And then, in each frame, for each frequency bin, the estimated LRSV is updated recursively with the maximum likelihood estimation of the current reverberant power. The effectiveness of the present method in terms of the quality of the recovered signals is confirmed by experiments.