Speech Dereverberation by Blind Adaptive MIMO Filtering Exploiting Nongaussianity, Nonwhiteness, and Nonstationarity

Inhalt:
In this paper, we present a class of novel algorithms for blind dereverberation of speech signals based on TRINICON, a general framework for broadband adaptive MIMO signal processing. In order to exploit all fundamental stochastic signal properties of speech for the dereverberation/deconvolution process and to avoid any whitening artifacts known from previous approaches, we propose the incorporation of a specially designed signal model based on an expansion using multivariate Chebyshev-Hermite polynomials. The multivariate model also inherently includes linear prediction which is known to be related directly to the human vocal tract model. The framework is applicable to both single-speaker scenarios and also to multiple simultaneously active speakers. In the latter case it also includes blind source separation in addition to the dereverberation.