Single channel speech enhancement based on prominent pitch estimation

A novel single channel speech enhancement algorithm based on prominent pitch estimation is proposed in this paper. In the traditional spectral substation method, the noise is assumed to be stationary. However our proposed method is based on the harmonic feature of speech signal which is not dependent on the environmental noise. The harmonic peaks for the objective speech can be extracted from the contaminated spectrum with the estimated prominent pitch. First the prominent pitch is estimated from the noisy speech based on Maximum-Harmonic-Energy and Minimum-Frequency-Deviation (MHEMFD), and then the harmonics are selected from the noisy spectral peaks according to harmonic model. Finally the enhanced speech signal is obtained by reconstructing the harmonic structures into the time domain based on the harmonic principle. The experimental results show that our algorithm outperforms some existed methods in terms of signal-to-noise ratio (SNR).