Monaural voiced speech separation based on harmonic and energy features

A monaural voiced speech separation approach based on harmonic and energy features is proposed. The method casts voiced speech separation as sound classification problem in time-frequency domain. First, the energy feature is employed to assist harmonic features. Then, the harmonic and energy features of time-frequency units with obvious harmonicity and large energy are replicated in the process of classifier training. Experimental results show that the proposed method obtains better signal-to-noise ratio improvement compared with previous approaches. The voiced speech separation is improved by introducing energy feature and feature replication.