A speech presence probability estimator based on fixed priors and a heavy-tailed speech model

Abstract

Speech enhancement approaches are often enhanced by speech presence probability (SPP) estimation. However, SPP estimators suffer from random fluctuations of the a posteriori signal-to-noise ratio (SNR). While there exist proposals that overcome the random fluctuations by basing the SPP framework on smoothed observations, these approaches do not take into account the super-Gaussian nature of speech signals. Thus, in this paper we define a framework that allows for modeling the likelihoods of speech presence for smoothed observations, while at the same time assuming super-Gaussian speech coefficients. The proposed approach is shown to outperform the reference approaches in terms of the amount of noise leakage and the amount of musical noise.