Abstract:

Pitch detection in a noisy environment plays an important role in speech signal processing. In order to effectively extract the pitch period in low signal to noise ratio (SNR), we propose a weighted auto correlation method based on wavelet packet transform. We employ the wavelet packet adaptive threshold to eliminate noise signals, and use the summation of approximate components after wavelet packet transform to emphasize the pitch information. Then we use the method of linear prediction error autocorrelation function weighted by wavelet packet coefficients to emphasize the peak of the true pitch period. Compared with the traditional methods based on autocorrelation function or wavelet-weighted, the experiments show that the proposed pitch extraction method has higher accuracy and smoother trajectory of pitch period. Moreover, it is robust when the SNR is -5dB. 

Key words: pitch detection, wavelet packet transform, linear prediction error；auto correlation function