Single-Channel Blind Dereverberation Based on Rank-1 Matrix Lifting in Time-Frequency Domain

Single-channel blind dereverberation aims to remove reverberation from a single-channel reverberant signal without using any additional information. In acoustics, weighted prediction error (WPE), which estimates a filter to predict the late reverberation and removes it, is widely used for this task. However, it is difficult to accurately estimate the prediction filter from a single-channel signal. In this paper, for better single-channel dereverberation, we propose to simultaneously estimates the source signal and the room impulse response (RIR) instead of only predicting reverberation. By modeling convolution using matrix lifting in the time-frequency domain, we formulate the dereverberation problem as a non-convex optimization problem of recovering a rank-1 matrix. The alternating direction method of multipliers (ADMM) is applied to approximately solve the optimization problem, resulting in closed form updates. In our experiments, we confirmed that the proposed method outperforms existing methods in several reverberant conditions.