Multi-Channel Linear Prediction Based on Binaural Coherence for Speech Dereverberation

Hong Liu, Xiuling Wang, Miao Sun, Cheng Pang


It has been shown that the multi-channel linear prediction (MCLP) can achieve blind speech dereverberation effectively. However, it always degrades the binaural cues which are exploited for human sound localization, i.e., interaural time differences (ITD) and interaural level differences (ILD). To overcome this problem, the multiple input-single output structure of conventional MCLP is modified to a binaural input-output structure for suppressing reverberation and preserving binaural cues simultaneously. First, by employing a binaural coherence model with head shadowing effects, the variance of desired signal can be estimated the same to both ears, which can ensure no modification of ILD. Then, the variance is utilized to calculate the prediction coefficients in a maximum-likelihood (ML) sense. Finally, the desired signals can be obtained as the prediction errors in MCLP. And since the algorithm does not disturb the phase of input signal, the ITD cue is kept. Evaluations with measured binaural room impulse responses (BRIRs) show that the proposed method yields a good performance on both speech dereverberation and binaural cues preservation.


DOI: 10.21437/Interspeech.2016-729

Cite as

Liu, H., Wang, X., Sun, M., Pang, C. (2016) Multi-Channel Linear Prediction Based on Binaural Coherence for Speech Dereverberation. Proc. Interspeech 2016, 1735-1739.

Bibtex
@inproceedings{Liu+2016,
author={Hong Liu and Xiuling Wang and Miao Sun and Cheng Pang},
title={Multi-Channel Linear Prediction Based on Binaural Coherence for Speech Dereverberation},
year=2016,
booktitle={Interspeech 2016},
doi={10.21437/Interspeech.2016-729},
url={http://dx.doi.org/10.21437/Interspeech.2016-729},
pages={1735--1739}
}