Further Analysis of the β-Order MMSE STSA Estimator for Speech Enhancement

In Bayesian approaches for speech enhancement, the clean speech is estimated by minimizing the expectation of a desired cost function. In the β-order MMSE STSA (βSA) Bayesian estimator, the cost function is the squared difference between the estimated and actual clean speech short-time spectral amplitude (STSA), both to the power β > 0. In this paper we propose an extension of the analysis of the βSA estimator for values of β < 0. We find that when β < 0, a normalization occurs in the βSA estimator which produces more noise reduction as β is reduced at the expense of additional speech distortion. Furthermore, the βSA estimator with β = -1 slightly outperforms the well known MMSE STSA and MMSE log-STSA (LSA) estimators in terms of the PESQ, for the two noises studied, while the overall MOS appreciation for β = -1 is found to be better than both MMSE STSA and LSA for white noise.