Multi-pitch Estimation for Speech Mixture Based on Multi-length Windows Harmonic Model

Harmonics of two concurrent speech signals may be overlapped in their mixing spectrum from the short time window analysis. We propose a multi-length windows (MLW) method based on harmonic model to estimate the multiple pitches of the single-channel mixed speech. The longest window is used to distinguish all the potential harmonic peaks and initially estimate the other pitch. And the other windows rectify the initial estimation. The advantages of our method include: i) it can obtain more accurate prominent pitch with autocorrelation; ii) it overcomes the low frequency resolution of the traditional harmonic model for estimating the other pitch. The simulation results show that the proposed algorithm outperforms the minimum mean square error (MMSE) method and short window (SW) harmonic method.