Research On Pitch Detection Algorithm Of Noisy Speech

In all fields of the digital processing of speech signal, for instance, speech analysis, speech synthesis, speech compression coding and speech recognition with confirm by speaker, etc, detecting pitch period accurately and reliably is extremely important, which will influence the performance of the whole system directly.Pitch detection methods, which are often used nowadays, can't get good performance in noise environment while it is inevitably influenced by background noise in the generation of real speech signal. So we focus on the study of pitch detection algorithms for noisy speech and strive to find algorithms with relatively good accuracy and robustness.In this paper, the generation of speech, its numeric model, its characteristics and its analyzing methods are introduced. The theory of wavelet analysis is also briefly expounded here. Besides, we make a classification, induction and summarization of the existing pitch detection methods. Finally, on the basis of the research of the theory of wavelet analysis and the pitch detection methods existing nowadays, some new pitch detection algorithms are proposed which are described as follows:(1) Based on the good denoising characteristic of wavelet transform, which can not only eliminate noise but also maintain the singularity characteristic of speech signal, we propose a pitch detection algorithm with the combination of wavelet transform and the second spectrum.(2) Based on the band pass characteristic of wavelet transform, we propose a pitch detection algorithm which combined wavelet transform with the signal waveform in time domain. Firstly, low frequency band signal including pitch information is extracted using wavelet transform and it is smoothed by a numerical filter to get a signal with stronger periodicity. Then, many features of the smoothed signal are integrated used here, for example, information of the width, area and magnitude of the peaks etc. Pitch period is detected by the matching degree of these information.(3) Based on the characteristic of multi-resolution analysis of wavelet transform, pitch detection algorithm with the combination of wavelet transform and weighted autocorrelation function is proposed here. In this algorithm, we weight and sum the approximate components of multi-level wavelet transform to emphasize the fundamental frequency information, and make it the base for pitch detection. Then pitch period is detected adopting the method of autocorrelation function weighted by the modified average magnitude difference function.(4) A new pitch detection algorithm is proposed using the theory of singularity detection of wavelet transform. Firstly, the speech signal is preprocessed to remove the influence of formant and noise, adopting linear predictive analysis with the combination of low pass filter whose cut-off frequency is 500Hz. Secondly, different methods are adopted to detect pitch periods of sonant frames and transition frames of surd and sonant. That's to say, the method of weighted autocorrelation function is adopted for sonant frames while the singularity detection theory of wavelet transform is adopted for transition frames of surd and sonant. Finally, individual errors in pitch track are corrected by a smoothing method, in which a searching and tentative method together with a judge mechanism of adding or eliminating pitch period is used.At last, programs of the proposed algorithms are designed and simulation experiments are carried out comparing with the traditional methods in different signal to noise ratio. Through experiments, it is shown that the proposed algorithms have high accuracy in pitch estimate, strong robustness to noise, so the performance improved evidently.