A sparsity-based method to solve the permutation indeterminacy in frequency domain convolutive blind source separation

Prasad Sudhakar (1) , Rémi Gribonval (1)

1 METISS - Speech and sound data modeling and processing

Existing methods for frequency-domain estimation of mixing filters in convolutive blind source separation (BSS) suffer from permutation and scaling indeterminacies in sub-bands. However, if the filters are assumed to be sparse in the time domain, it is shown in this paper that the l1-norm of the filter matrix increases as the sub-band coefficients are permuted. With this motivation, an algorithm is then presented which solves the source permutation indeterminacy, provided there is no scaling indeterminacy in sub-bands. The robustness of the algorithm to noise is also presented.

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Prasad Sudhakar, Rémi Gribonval. A sparsity-based method to solve the permutation indeterminacy in frequency domain convolutive blind source separation. ICA 2009, 8th International Conference on Independent Component Analysis and Signal Separation, Mar 2009, Paraty, Brazil. ⟨10.1007/978-3-642-00599-2_43⟩. ⟨inria-00544760⟩