Separation of speech signals under reverberant conditions

BSS performance is not still enough for speech signals and long acoustic responses. An original frequency model, strictly equivalent to a time linear convolution, is used for speech signals under highly reverberant conditions. If the responses are virtually sectioned in K blocks of N samples, the time linear convolutions are strictly transformed in frequency domain at frequency ν, into FIR filtering of K taps where the K taps are the complex gains of the K sectioned blocks at the same frequency ν. Short values of the DFT, N, can be employed, although the length of the responses remains long enough (K.N samples) to suit with acoustic responses. Finally, the separation is achieved with a natural gradient algorithm based on a maximum-entropy cost function. The proposed method is then tested on speech signals.