Single‐Channel Speech Separation Based on Non‐negative Matrix Factorization and Factorial Conditional Random Field. Issue 5 (1st September 2018)

• Record Type: Journal Article
• Authors: LI, Xu; TU, Ming; WANG, Xiaofei; WU, Chao; FU, Qiang; YAN, Yonghong
• Is Part Of: Chinese journal of electronics. Volume 27:Issue 5(2018)
• Page Start: 1063
• Page End: 1070
• Publication Date: 2018-09-01
• Subjects: Single‐channel speech separation -- Nonnegative matrix factorization -- Factorial conditional random field -- k‐means clustering; hidden Markov models -- matrix decomposition -- source separation -- speech processing; nonnegative matrix factorization based algorithm -- sparse NMF -- spectral structure -- k‐means clustering -- FCRF model -- factorial conditional random field model -- standard NMF -- Markov model -- temporal dynamics -- state sequence -- dictionaries -- speech signal -- temporal continuity -- prior known speakers -- single‐channel speech separation -- noise figure 1.12 dB -- noise figure 2.37 dB -- noise figure 0.4 dB -- noise figure 2.04 dB -- noise figure 4.26 dB -- noise figure 0.62 dB -- noise figure 2.8 dB -- noise figure 5.08 dB -- noise figure 1.06 dB; Electronics -- Periodicals; Electronics -- China -- Periodicals; Electronics; China; Periodicals; 621.38105
• DOI: 10.1049/cje.2018.06.016
• Languages: English
• ISSNs: 1022-4653

Abstract

Abstract : A new Non‐negative matrix factorization (NMF) based algorithm is proposed for single‐channel speech separation with a prior known speakers, which aims to better model the spectral structure and temporal continuity of speech signal. First, NMF and k ‐means clus‐ tering are employed to obtain multiple small dictionaries as well as a state sequence that describes the temporal dynamics between these dictionaries for each speaker. Then, a Factorial conditional random field (FCRF) model is trained using the state sequences and dictionaries to jointly model the temporal continuity of two speakers' mixed signal for separation. Experiments show that the proposed algorithm outperforms the baselines with respect to all metrics, for example sparse NMF (+1.12dB SDR, +2.37dB SIR, +0.40dB SAR, +0.2 MOS), nonnegative factorial hidden Markov model (+2.04dB SDR, +4.26dB SIR, +0.62dB SAR, +1.0 MOS) and standard NMF (+2.8dB SDR, +5.08dB SIR, +1.06dB SAR, +1.2 MOS).