IMPROVED SPEECH SPECTRAL VARIANCE ESTIMATION UNDER THE GENERALIZED GAMMA DISTRIBUTION

DFT-based single-microphone speech enhancement methods need an estimate of the clean speech spectral variance. Often the ”decision-directed” spectral variance estimato r is used, because of its good performance: it strongly reduces the musical noise phenomenon. It has recently been shown that this estimator is severely biased at low SNRs when the smoothing factor approaches one [1]. Here we propose a variance estimator with reduced bias, under the assumption of a generalized Gamma distribution for the clean speech spectral amplitudes. For the reconstruction, the MMSE estimator of the amplitudes themselves is used, derived under the same distribution assumption. For the same speech quality versus noise reduction trade-off, the new variance estimator leads to less musicality.