Abstract
In a real environment, sound sources are mixed through convolution mixture, and it is difficult to separate sources in the time domain. Therefore, we use independent component analysis (ICA) in the frequency domain. Using ICA in the frequency domain could reduce the computation, but there are two important ambiguities: scaling problem and permutation problem. These ambiguities affect reconstruction of separated source. In this thesis, a new approach is proposed for solving the scaling problem and permutation problem. Besides, Time difference of arrival (TDOA) is used to confirm that two sources exist simultaneously. To solve the scaling problem, the Gaussian mixture model is uesd to approximate the distribution of the separated signal and the mixed signal. The difference between the mean of separated signal and the mean of the mixed signal is compensated to solve the scaling problem. Considering the permutation problem, the present algorithm relies on the assumption that the correlations should be high between the temporal envelopes of neighboring frequencies from the same sound source. First, we find the five neighboring frequency bins which have a high correlation with each other as a standard. After that, separated source in other frequency bins could confirm permutation through the correlation with the standard. We compare with the result of the approach of [30]. Computation time is reduced by 17 seconds and SIR enhances by 4 dB. In the part of the questionnaire, we get a higher score than [30]. 66 subjects were recruited to conduct a listening comprehension test. The accuracy of listening comprehension of separated sources is 41%, 26%, 45% higher than unprocessed sounds. The results show that our approach reduces computation cost and enhances sound quality when compared to the existing method [30].