Pitch Detection Based On Multitaper Spectrum And Mean Shift Algorithm

With the development of modern science and technology, speech signal processing technology is very mature. Pitch is an important parameter of speech characteristics. However, the speech is often accompanied by noise. Noise can interfere with the accuracy of the speech parameter extraction. So pitch detection from noisy speech is an important and difficult subject.This paper studies pitch and voiced/ unvoiced detection from the noisy multitaper spectrum with mean shift algorithm. Firstly, the performance of the multitaper spectrum is studied. It can be resulted that multitaper spectrum can smooth noise. Secondly, It is a time-frequency domain algorithm. An apposite initial pitch is set and the higher harmonics are iterated to make mean move towards the pitch frequency. In this way, the pitch can be detected reliably. Finally, the mean shift combines with multitaper windows and pitch and voiced/unvoiced can be detected in low SNR.It can be concluded that the performance of voiced/ unvoiced detection in low SNR is improved by multitaper spectrum. The performance of Sine windows is better than the Slepian windows. Compared to other algorithms, the mean shift algorithm has less code and runs more fast. In addition the noise robustness is better.