SYSTEM FOR THE BLIND SEPARATION OF ACOUSTICAL SIGNALS FROM CONVOLUTIVE MIXTURES

摘要

The invention relates to a system for the blind separation of acoustical signals from convolutive mixtures, without knowing data about the initial sources. According to the invention, the system comprises an analysis block (2) comprising a number of l identical analysis banks, each having a set of l band-pass filters (5) which decompose a mixture signal on l sub-bands, a neural computing block (3) formed of neural network banks (6) which separate the sources signals from the mixture on each sub-band and a synthesis block (4) consisting of filter banks (7) which reconstruct the initial sources signals on the basis of the outputs of the neural computing block (3).