Underdetermined Speech Blind Source Separation based on Harmonics Extraction

The existing underdetermined speech blind source separation (BSS) methods can hardly concurrently possess high efficiency and high performance. To solve this problem, a harmonics extraction based underdetermined speech BSS algorithm is proposed in this paper. Firstly, introduce the spectrum correction technique to extract the harmonic components from the mixtures’ short time Fourier transform (STFT); Secondly, apply a phase-coherence criterion on these harmonic components to identify the single source components; Thirdly, employ the adaptive k-means clustering on these refined single source patterns to estimate the source number and the mixing matrix; Finally, combining the subspace projection algorithm with this estimated matrix yields the source recovery. Specifically, the combination of harmonics extraction and single -source component identification ensures that the mixing matrix can be accurately estimated in low complexity. Simulation demonstrates that, compared to the original subspace projection algorithm, the proposed BSS method can acquire a higher recovery quality, which presents a potential application in other harmonic related fields.