Blind Dereverberation of Audio Signals

This project examines the problem of single channel blind dereverberation. After estimating the T60 value, a time-domain binary masking approach was used to remove regions of the signal that were largely dominated by reverberant energy. Performance of the system was examined for several different classes of audio (hand clapping, drums, and speech) and for varying amounts of reverberation. In addition to subjective evaluation via spectrograms and audio output, results were assessed by comparing the accuracy of the binary mask to an ideal binary mask calculated using a clean reference signal. Results were generally good although as expected, performance was better for sparse signals. Interestingly, some mask accuracy results were better when more reverberation was present, even though subjectively the system performed worse on these examples.