Speech enhancement combined with dereverberation and acoustic echo reduction for time varying systems

This paper deals with speech enhancement problems for highly time-varying systems. For high noise-reduction performance with little speech distortion, cascade methods of various linear/non-linear filters can be easily conceived of, but cascade methods are partial optimizations and cause interferences between filters, because an optimized output signal of a preceding filter is not always an optimized input signal of successive filters. To obtain total optimization, the proposed method adapts all filters under a unified likelihood function with an extended local Gaussian model, which consists of non-stationary multichannel speech model with time-varying acoustic transfer functions, an acoustic echo model with a time-varying echo path model, and stationary background noise model. Qualitatively, the residual reverberation or the residual acoustic echo after the preceding linear filtering keep desirable time-structures for the successive non-linear filtering. Experimental results show that the propose method is superior to the cascade methods.