We propose a multi-channel speech dereverberation approach based on cross-channel cancellation and spectrogram decomposition. The reverberation is modeled as a convolution operation in the spectrogram magnitude domain. Using the Itakura divergence (I-divergence), we decompose reverberant spectrogram into clean spectrogram convolved with a deconvolution filter. The speech spectrogram is constrained and regularized by non-negativity and sparsity, respectively, while the deconvolution filter is constrained by non-negativity and cross-channel cancellation. Spectrogram decomposition of individual channels and cross-channel cancellation are jointly optimized by a multiplicative algorithm to achieve multi-channel speech dereverberation. We confirm through experiment that the proposed multi-channel method outperforms the related other single/multi-channel methods.
Index Terms: Multichannel dereverberation, Spectral decomposition, I-divergence, Sparsity, Cross-channel cancellation
Experiment 1: Demo of the Proposed Method
200 ms Reverberant Processed
400 ms Reverberant Processed
600 ms Reverberant Processed
800 ms Reverberant Processed
room Reverberant Processed
Experiment 2: Comparison (see explanation)
Reverberant (a) (b) (c) (d) (e) (f) (g)
Bibliographic reference. Yu, Meng / Soong, Frank K. (2012): "Constrained multichannel speech dereverberation", In INTERSPEECH-2012, 1938-1941.