FAST INDEPENDENT VECTOR ANALYSIS USING NON-OVERLAPPING FREQUENCY SUBBANDS PARTITION AND POWER RATIO CORRELATION

摘要

The blind source separation of convolutive mixtures in frequency domain needs resolving the well-known permutation ambiguity problem. Traditional independent vector analysis (IVA) algorithms solve this problem using higher-order dependency features of the input signals, where some of them implement in the whole frequency range and the others in overlapping cliques or subbands. Whereas, the overlapping schemes cannot always lead to the correct permutation in reverberant and noisy environment. This paper proposes a novel fast IVA algorithm using non-overlapping frequency subbands partition and power ratio correlation to align the permutation. The separation performance of the proposed algorithm is evaluated using measured room impulse responses and its robustness is also confirmed by adding different levels of white Gaussian noise into the mixtures.