The Research Of Pitch Detection Methods In Speech Signal

Language is the most important tool of communication.It is the dream of peoplefor a long time to communicate with machine and let the machine to understand whatwe say. With the development of technology, speech signal processing will becomeindispensable as an important component in the information war.In speech signal processing, the key is to analyze the essential characteristics ofthe speech signal.Only get a better grasp in the detection of characteristic parameters,can we use these parameters to realise speech enhancement, speech recognition,speech synthesis and voice compression coding and other processing. Pitch period isone of the important parameters to describe the excitation source. Based on thecurrent research status of pitch detection, this paper made a research work as follows:(1) Through analyzing the traditional autocorrelation method, we find that thenumber of the coefficients for sums reduced with the increase of the time lag, whichleads the peak of the autocorrelation function decreases gradually and bringsdifficulties for peak detection. According to the situation, an improved autocorrelationfunction is proposed. In order to overcome the peak decrease with the increase of thetime lag, amplitude adjustment is added in the function. Experiments show that thenew algorithm overcomes the amplitude attenuation and highlight the peakinformation.(2) Autocorrelation function method and wavelet transform are the classical pitchdetection methods. Autocorrelation function exist the impact of double frequencyerrors and half frequency errors in pitch detection, wavelet exist the impact of pseudorandom point errors in pitch detection. The shortcomings of these methods areanalyzed when they are used alone, on the basis of this, a new method combiningimproved autocorrelation function with weighted wavelet components is proposed.The components of multilevel wavelet transform are weighted to emphasize thefundamental frequency information. And then the two methods are combined toemphasize the peak at the position of the true pitch period. Experiments show thatcompared with the classical methods, the method can decrease double frequencyerrors、half frequency errors and pseudo random point errors, the precision of the pitchdetection is improved. (3) According to the non-stationary and nonlinear time-varying characteristics ofspeech signal, a speech pitch detection method based on Hilbert-Huang transform ispresented in this paper. Firstly, use the short-time energy to judge voice and unvoice,then the signal is decomposed into a number of intrinsic mode functions,with Hilberttransform, the instantaneous frequency and instantaneous amplitude of each intrinsicmode function are obtained,the components of the intrinsic mode functions areweighted to emphasize the fundamental frequency information according to thecharacteristics of pitch, lastly, the square of the autocorrelation is used to detect thepitch. Experiments show that compared with the classical methods, the proposedmethod provides a higher precision and better robustness.