Blind speech extraction using subband independent component analysis with scale adjustment function and neural memory

Takeshi Hanada, Tetsuya Hoya, Takahiro Murakami, Yoshihisa Ishida

Abstract: This paper presents a new method for speech extraction by a combined subband independent component analysis (subband ICA) and neural memory. In the method, subband ICA separates the signals obtained from microphones into the signal components of interest and the rest at each subband. Subband ICA approach has two fundamental problems. The neural memory represented by probabilistic neural networks (PNNs) is then used for identifying the signal components of interest among the separated components and thereby solved the permutation problem. Then we adjust the scale of each output signal to that of the corresponding input signal. Simulation results for both the instantaneous and delayed mixture cases show that the proposed subband ICA approach consistently yields the performance improvement in comparison with the conventional fullband/subband approaches.