Blind Separation of Uncorrelated Sound Sources from the Principle of Least Spatial Complexity

Bin Dong (1) , Jérôme Antoni (1)

1 LVA - Laboratoire Vibrations Acoustique

A novel algorithm based on the principle of least spatial complexity is proposed to decom-pose a sound field into uncorrelated sources by minimizing its spatial entropy. The global sound source distribution is first reconstructed from measurements returned by an array of microphones. Then it is transformed into uncorrelated virtual sources by means of the eigen-value decomposition. However, it is well-known that having uncorrelated sources is a neces-sary but by no means a sufficient condition for separation. In order to find a unique solution with physical meaning, it is proposed to invoke the principle of least spatial complexity which forces an optimal solution with maximum spatial compactness. This is accomplished thanks to a novel conjugate gradient algorithm operating on the Stiefel manifold. An experi-ment is conducted to validate the proposed algorithm and evaluate its performance. Final re-sults show that the sound field of each separated source perfectly matches that of the actual one, as would be obtained if all other sources could be switched off one by one.

Domaines: Acoustique [physics.class-ph]; Acoustique [physics.class-ph]

Bin Dong, Jérôme Antoni. Blind Separation of Uncorrelated Sound Sources from the Principle of Least Spatial Complexity. 19th International Congress on Sound and Vibration, Jul 2012, Lithuania. ⟨hal-01019706⟩