Models for blind speech dereverberation: A subband all-pole filtered block stationary autoregressive process

Single channel blind dereverberation of speech acquired in an acoustic environment is approached using parametric modelling and estimation theory to obtain channel estimates; inverse filtering is then applied to derevererate the speech. Models previously used in such an approach deal with relatively simple scenarios such as a gramophone horn modelled with 70 parameters; the weakness of those models, however, is in their attempt to simultaneously model the full channel spectrum by a single all-pole filter. Not only does this lead to a large computational load, it is not parsimonious, nor is it scalable such that the algorithm can be applied to higher dimensional problems. A better approach uses subbands; in this paper, a subband all-pole filter models the channel while the source is still represented by a single-band block stationary AR process. An example is given of blindly dereverberating a signal observed through the aforementioned gramophone horn, demonstrating an equally robust, but more flexible and scalable model.