Single-Channel Speech Dereverberation Based on Non-negative Blind Deconvolution and Prior Imposition on Speech and Filter

In this letter, we describe a single-channel speech dereverberation method in the short-time Fourier transform domain by using non-negative blind deconvolution. Robust decomposition of the magnitude spectra of the reverberated speech into its clean speech and a reverberation filter can be achieved by imposing a sparse and frequency-dependent prior model on the speech and an exponentially decaying envelope on the filter. Subsequently improved dereverberated speech is estimated without crude speech prior imposition for the fixed reverberation filter. The effectiveness of the algorithm was demonstrated with experimental results on speech reverberated by room impulse responses.