Speech separation microphone array based on law of causality and frequency domain processing

This paper proposes a microphone array system separating speech signals, based on a different principle from independent component analysis (ICA). This system applies linear prediction error filters to microphone outputs, and using the prediction errors, adjusts the coefficients of adaptive filters. In this case, only the prediction errors satisfying the law of causality become available for the adjustment; consequently, this system can steer a null toward a direction satisfying it. The permutation problem discussed in ICA can be thereby avoided. In practical use, the speech signals are required to be separated as quickly as possible. This paper proposes applying a frequency domain technique to the adjustment of the coefficients of adaptive filters. The application of the frequency domain technique is expected to improve the adjustment speed. Computer simulations show that the adjustment time can be reduced to about two second and that the separation effect of 30 dB can be obtained.