Multichannel sub-band inverse filtering for room dereverberation

Abstract

Room dereverberation is difficult because the impulse responses are long, nonminimum phase and change significantly with time and source/receiver positions. The problem was approached in two steps: (1) Explore the possibility of inverse filtering when a training signal is available; (2) blindly estimate impulse responses from reverberant signals and then perform inverse filtering. It is found that by dividing the signals into a large number of sub-bands, the impulse responses of many of the sub-bands become invertible. The invertibility of each sub-band depends on the zero distribution of the full band impulse response. When multiple microphones are used, full band dereverberation is achieved by selecting the best microphone in each sub-band and reconstructing the full band signal from dereverberated sub-band signals. An almost perfectly recovered reverberant speech of RT=0.85 s, when the training signal is used, will be demonstrated. For the second step, reverberant signals from multiple sensors are divided into sub-bands. In each sub-band, the ratio of impulse responses from two sensors is estimated by the least-mean-square estimation. Pade approximation is used to estimate each individual impulse response. The sub-band signals are then inverse-filtered and combined to form a full band dereverberated signal.