Monaural voiced speech segregation based on elaborate harmonic grouping strategies

In this paper, an enhanced algorithm based on several elaborate harmonic grouping strategies for monaural voiced speech segregation is proposed. Main achievements of the proposed algorithm lie in three aspects. Firstly, the algorithm classifies the time-frequency (T-F) units into resolved and unresolved ones by carrier-to-envelope energy ratio, which leads to more accurate classification results than by cross-channel correlation. Secondly, resolved T-F units are grouped together according to minimum amplitude principle, which has been verified to exist in human perception, as well as the harmonic principle. Finally, "enhanced" envelope autocorrelation function is employed to detect amplitude modulation rates, which helps a lot in reducing half-frequency error in grouping of unresolved units. Systematic evaluation and comparison show that performance of separation is greatly improved by the proposed algorithm. Specifically, signal-to-noise ratio (SNR) is improved by 0.96 dB compared with that of previous method. Besides, our algorithm is also effective in improving the PESQ score and subjective perception score.