A Neural Network Blind Separation Method Based on Special Frequency Bins

Abstract

It is a usual approach to separate the convolutive mixtures blindly in frequency domain. Many blind separation algorithms proposed for instantaneous mixtures were employed to separate signals in each frequency bin. These approaches must consider all frequencies, and correct the permutation/amplitude of output signals, resulting in a huge computation. In this paper we propose a neural network blind separation approach with a few special frequency bins, which have line spectra or some foreknowing characteristics. The approach can separate convolutive signals effectively with a reduced computation, suitable for the application on real time. The validity and performance of the proposed approach is demonstrated by computer simulation with speech and ship radiating underwater acoustic signals. The comparison between the proposed method and all frequency bins algorithms is performed on their calculation complexity and separation effect.