Reverberation corrupts speech recorded using distant microphones, resulting in poor speech intelligibility. We propose a single-channel, supervised non-negative matrix factorization (NMF) based dereverberation method, in contrast to the convolutive NMF (CNMF) based methods in literature. Recent supervised approaches use a CNMF model for reverberation and a NMF model for clean speech spectrogram to obtain enhanced speech by directly estimating the clean speech activations. In the proposed method, with a separability assumption on the room impulse response (RIR) spectrogram, the reverb speech can be decomposed into bases and activations using conventional NMF. Using these reverb activations, the clean speech activations are estimated to obtain enhanced speech. The proposed model (i) helps in imposing meaningful constraints on the RIR in both frequency- and time-domains to achieve improved enhancement (ii) leads to a framework that can include a NMF model for noise. (iii) gives a better interpretation of the effects of reverberation in the NMF context. We evaluate and compare the enhancement performance of the algorithm on reverb and noisy conditions, simulated using TIMIT utterances and REVERB challenge RIRs. The proposed method performs better than existing C-NMF based methods in objective measures, such as cepstral distance (CD) and speech-to-reverberation modulation energy ratio (SRMR).
Cite as: Mohanan, N., Velmurugan, R., Rao, P. (2018) A Non-convolutive NMF Model for Speech Dereverberation. Proc. Interspeech 2018, 1324-1328, doi: 10.21437/Interspeech.2018-1834
@inproceedings{mohanan18_interspeech,
author={Nikhil Mohanan and Rajbabu Velmurugan and Preeti Rao},
title={{A Non-convolutive NMF Model for Speech Dereverberation}},
year=2018,
booktitle={Proc. Interspeech 2018},
pages={1324--1328},
doi={10.21437/Interspeech.2018-1834}
}