A New Method Based on Spectral Subtraction for Speech Dereverberation

TLDR

Reverberation suppression differs from classical speech denoising because the reverberation noise is non‑stationary. The paper proposes a monaural spectral‑subtraction method that estimates the non‑stationary reverberation‑noise power spectrum using a statistical model of late reverberation. The method applies spectral subtraction to estimate and suppress late reverberation, and is evaluated on real reverberated speech recordings. Across room‑impulse‑response lengths of 0.34 s to 1.7 s, the method achieves significant noise reduction with minimal distortion, and markedly improves ASR accuracy in diverse reverberant settings.

Abstract

Summary A new monaural method for the suppression of late room reverberation from speech signals, based on spectral subtraction, is presented. The problem of reverberation suppression differs from classical speech de-noising in that the “reverberation noise” is non stationary. In this paper, the use of a novel estimator of the non-stationary reverberationnoise power spectrum, based on a statistical model of late reverberation, is presented. The algorithm is tested on real reverberated signals. The performances for different RIRs with ranging from 0.34 s to 1.7 s consistently show significant noise reduction with little signal distortion. Moreover, when used as a front end to an automatic speech recognition system, the algorithm brings about dramatic improvements in terms of automatic speech recognition scores in various reverberant environments.