Two-stage dereverberation with integrated reverberation and noise estimation

This paper proposes a two-stage dereverberation algorithm with integrated reverberation and noise estimation. In the first stage, an inverse filter to the room impulse response suppresses reverberation by early reflections. Minimum mean-square error short-time spectral amplitude estimator in the second stage minimizes influence of both late reflections and ambient noise, simultaneously. Due to this integrated structure, a postprocessing noise suppressor can be eliminated. Evaluations with different room impulse responses show that the maximum ACR (absolute category rating) score improvement over a conventional algorithm exceeds 1.0. It is also shown that a CCR (comparison category rating) score of as much as 1.3 is obtained compared to the conventional algorithms.