Co-channel speech separation based on sinusoidal model for speech

This paper proposed a sinusoidal model based cochannel speech separation method which is based on the assumption that a speech wave consists of a series of sinusoidal signals with time-varying amplitudes, frequencies and phases. In this paper subharmonic summation was expanded to overlapping speech situations to detect and separate the pitch frequencies of each speech. Then 2-D harmonic sieve arithmetic was proposed to obtain the harmonic frequencies, amplitudes and phases of each speech. Finally the components of each speech were separated by a sinusoidal model. This method is mainly suitable for overlapping vowels, and the experiment results showed it can effectively separate two overlapping speeches even when the pitch frequencies of the two speeches are relatively close