We have developed an on-line separation method for audio signals. The adopted approach makes use of the time-frequency transform of the signals as a sparse decomposition. Since the sources for the most part do not overlap in the time-frequency domain, we get raw estimates of their individual mixing parameters with an analysis of the mixture ratios. We then obtain reliable mixing parameters by dynamically clustering these instantaneous estimates. The mixing parameters are used to separate the mixtures, even at time-frequency points where the sources overlap. In addition, even when the mixing parameters change over time, our approach is able to separate signals with only one pass through the data. We have evaluated this approach first on computer generated anechoic mixtures and then on real echoic mixtures recorded in a car.
Cite as: Bourgeois, J. (2003) A clustering approach to on-line audio source separation. Proc. 8th European Conference on Speech Communication and Technology (Eurospeech 2003), 1745-1748, doi: 10.21437/Eurospeech.2003-29
@inproceedings{bourgeois03_eurospeech,
author={Julien Bourgeois},
title={{A clustering approach to on-line audio source separation}},
year=2003,
booktitle={Proc. 8th European Conference on Speech Communication and Technology (Eurospeech 2003)},
pages={1745--1748},
doi={10.21437/Eurospeech.2003-29}
}