Hybrid Approach to Single-Channel Speech Separation Based on Coherent-Incoherent Modulation Filtering

Single-channel speech separation is a challenging problem that has been of particular interest in recent years. Here the goal is to separate the target speech signal from the interference signals, with high accuracy. We propose a new hybrid single-channel speech separation system that applies adaptive coherent modulation filtering for low-frequency subbands and iterative incoherent speech separation technique for high-frequency subbands. In the adaptive coherent modulation filtering, an affine projection filter is applied to subband envelope in order to eliminate the interference signal. The subband envelope is determined via demodulation of the subband signal using a coherently detected subband carrier based on the time-dependent spectral center-of-gravity demodulation. The adaptive affine projection filter uses the separated target signal obtained from the iterative incoherent speech separation system as a reference signal. This system first obtains a rough estimate of target fundamental frequency range and then uses this estimate to segregate target speech. It then improves both fundamental frequency range estimation and voiced speech separation iteratively. Perceptual evaluation of speech quality, as one of the evaluation indices investigated in this paper, indicates that the proposed system extracts the majority of target speech segments with minimal interference and outperforms previous systems in voiced speech separation.