Efficient dereverberation in networked audio systems

摘要

Features are disclosed for performing efficient dereverberation of speech signals captured with single- and multi-channel sensors in networked audio systems. Such features could be used in applications requiring automatic recognition of speech captured with sensors. Dereverberation is performed in the sub-band domain, and hence provides improved dereverberation performance in terms of signal quality, algorithmic delay, computational efficiency, and speed of convergence.

主权项

1. A device for reducing reverberation in an audio signal, the device comprising:
computer-readable memory storing executable instructions; one or more physical computer processors in communication with the computer-readable memory, wherein the one or more physical computer processors are programmed by the executable instructions to at least:
receive an input audio signal;determine a first sub-band sample from the input audio signal, wherein the first sub-band sample corresponds to a first capture time range and a first frequency band, the first capture time range identifying a first period of time during which the first sub-band sample was captured;obtain first dereverberation weights corresponding to the first frequency band;determine a first dereverberated sub-band sample using the first dereverberation weights, the first sub-band sample, and a first plurality of sub-band samples corresponding to a period of time of capture preceding the first capture time range, wherein the first dereverberated sub-band sample corresponds to the first frequency band and the first capture time range, and the first plurality of sub-band samples includes samples having frequencies included in the first frequency band;generate a first dereverberated output audio sample using the first dereverberated sub-band sample;determine a second sub-band sample from the input audio signal, wherein the second sub-band sample corresponds to a second capture time range and the first frequency band, the second capture time range identifying a second period of time of capture occurring after the first capture time range;obtain a first Cholesky factor of a first matrix corresponding to the first dereverberation weights;generate a second Cholesky factor of a second matrix using the second sub-band sample, the first Cholesky factor, and a second plurality of sub-band samples corresponding to a third period of time of capture preceding the second capture time range and including samples having frequencies included in the first frequency band;generate second dereverberation weights using the second Cholesky factor;generate a second dereverberated sub-band sample using the second dereverberation weights, the second sub-band sample, and the second plurality of sub-band samples, wherein the second dereverberated sub-band sample corresponds to the first frequency band and the second capture time range; andgenerate a second dereverberated output audio sample using the second dereverberated sub-band sample.