Modulation spectrum based beamforming for speech enhancement
Sam Karimian-Azari, Tiago Henrique Falk · 2017
In array signal processing, beamforming is a common technique to align time differences between multi-microphone signals. Beamformers, however, have limits to reduce noise specially in the presence of reverberation. In this paper, we incorporate modulation properties of speech into a pre-processing algorithm to improve beamformer performance under combined noise-plus-reverberation conditions. In the modulation domain, signals are decomposed into modulators and carriers. Here, we propose to filter and perform short-time spectral subtraction of the modulator as a pre-processing step prior to beamforming, which in turn, is designed to align time differences between carriers of the array signal and to minimize the residual noise of the pre-processed signals. Simulation results with several noise-only and noise-plus-reverberation conditions show that the modulation pre-processing has improved the minimum power distortionless response beamformer by up to 7.4dB in the signal-to-noise ratio and 0.6 points in perceptual evaluation of speech quality (PESQ) score.