BSS Sound Demonstration

We have been working on blind source separation (BSS) for convolutive mixtures of speeches. We employ frequency-domain BSS where independent component analysis (ICA) is applied separately in each frequency bin. The key technique in the frequency-domain BSS is how to solve the permutation problem. We have proposed a good method for solving the permutation problem [1,2], which is applicable even for more than two sources [3]. Also, we have proposed a method to mitigate the circularity effect of frequency-domain BSS [4].

Experimental setups

Sound files

2 sources3 sources4 sources
Source signals a   b a   b   c a   b   c   d
Observed signals x1   x2 x1   x2   x3 x1   x2   x3   x4
Decomposition of
observed signals
(xj = Σi xji)		 x11   x12
x21   x22 		x11   x12   x13
x21   x22   x23
x31   x32   x33 		x11   x12   x13   x14
x21   x22   x23   x24
x31   x32   x33   x34
x41   x42   x43   x44
Separated signals y1   y2 y1   y2   y3 y1   y2   y3   y4
Average SIR (dB) 17.34 12.49 9.26

References

  1. H. Sawada, R. Mukai, S. Araki, S. Makino, "A Robust and Precise Method for Solving the Permutation Problem of Frequency-Domain Blind Source Separation," 4th International Symposium on Independent Component Analysis and Blind Signal Separation (ICA2003), pp. 505-510, Apr. 2003. (Paper: PDF)

  2. H. Sawada, R. Mukai, S. Araki, S. Makino, "A Robust Approach to the Permutation Problem of Frequency-Domain Blind Source Separation," IEEE International Conference on Acoustics, Speech, and Signal (ICASSP2003), pp. 381-384, Apr. 2003.

  3. H. Sawada, R. Mukai, S. Araki, S. Makino, "Convolutive Blind Source Separation for more than Two Sources in the Frequency Domain," IEEE International Conference on Acoustics, Speech, and Signal (ICASSP2004), accepted.

  4. H. Sawada, R. Mukai, S. de la Kethulle, S. Araki, S. Makino, "Spectral Smoothing for Frequency-Domain Blind Source Separation," International Workshop on Acoustic Echo and Noise Control (IWAENC2003), pp. 311-314, Sep. 2003. (Paper: PDF)