A Robust Blind Source Separation Algorithm Based on Non-negative Matrix Factorization and Frequency-Sliding Generalized Cross-Correlation

Abstract

In this paper, a blind source separation algorithm based on time delay estimation (TDE) and non-negative matrix factorization (NMF) is proposed. In the TDE module, sub-band generalized cross correlation (GCC), frequency-sliding and singular value decomposition (SVD) techniques are used to get more accurate estimation. The time differences of arrival (TDOA) estimation of sources can be obtained from weighted low-rank approximation of the Frequency-Sliding GCC (FS-GCC) matrix. Then the sound sources are reconstructed in the NMF module by grouping the dictionary atoms according to their spatial information. The experiment uses utterances from SiSEC2018 database. Performance is quantified using the BSS Eval toolkits. Results prove that the proposed algorithm outperforms the compared ones in the noisy environments.