Binaural dereverberation based on delayed widely linear prediction in the time domain

Sound spatial information benefits human listeners in reverberant environments. This paper deals with the problem of binaural dereverberation, which reduces reverberation and meanwhile preserves the sound spatial information at the binaural outputs. A widely linear (WL) filtering framework is adopted where the multiple real microphone signals are merged into complex signals. The desired binaural outputs are also converted into complex signals with one channel being its real part, and the other channel being its imaginary part. By doing so, we transform the problem of binaural dereverberation to one of monaural dereverberation. In such a framework, the complex late reverberation is modeled using the multichannel delayed WL prediction by fully taking advantage of the noncircularity of the complex signals. A maximum likelihood method is then developed to estimate the optimal prediction filter with the speech signal of interest being modeled by a complex normal distribution. The relationship between the proposed method and the weighted prediction error (WPE) method is also discussed. Finally, simulation results are provided to justify the effectiveness of the developed method.