Parametric modelling for single-channel blind dereverberation of speech from a moving speaker

Single-channel blind dereverberation for the enhancement of speech acquired in acoustic environments is essential in applications where microphone arrays prove impractical. In many scenarios, the source-sensor geometry is not varying rapidly, but in most applications the geometry is subject to change, for example when a user wishes to move around a room. A previous model-based approach to blind dereverberation by representing the channel as a linear time-varying all-pole filter is extended, in which the parameters of the filter are modelled as a linear combination of known basis functions with unknown weightings. Moreover, an improved block-based time-varying autoregressive model is proposed for the speech signal, which aims to reflect the underlying signal statistics more accurately on both a local and global level. Given these parametric models, their coefficients are estimated using Bayesian inference, so that the channel estimate can then be used for dereverberation. An in-depth discussion is also presented about the applicability of these models to real speech and a real acoustic environment. Results are presented to demonstrate the performance of the Bayesian inference algorithms.