Pitch Detection Algorithm In Low Snr Research

In all fields of the digital processing of speech signals, for instance,speech analysis, speech synthesis, speech compression coding, andspeech recognition confirmed by speaker, etc, detecting pitch periodaccurately and reliably is an essential task. This paper aims at looking fora kind of pitch detection algorithm which can still obtain pitch period inlow Signal to Noise Ration (SNR). On the bases of classicalwavelet-based pitch detection, a new pitch period detection method basedon lifting wavelet transform is proposed combining Teager energyoperator (TEO) with spatial correlation function. The principle researchwork and achievements of the paper are shown as follows:(1) An improved pitch detection method based on the traditionautocorrelated function(ACF) is proposed. Using efficient pre-processingsuch as lowpass and numerical value filter before the process of the pithdetection; after the process of the pith detection, using search sound outsmoothness method. Experiments show that this algorithm has a betterrobustness and greater precision of pitch detection compared with theclassical ACF under the general background noise.(2)The use of lifting wavelet transform in pitch detection is studied.The wavelet transform can analysis the oddity position of signal betterthan the other methods. Firstly, the paper analyses the feasibility andadvantage of wavelet transform on pitch detecting, then using of liftingwavelet transform in pitch detection. Compared with the classic waveletalgorithm, the amount of computation is decreased to a half. Thealgorithm realizes the original position computation, and no excessmemory is needed in the computation process.(3)On the detect of voiced regions subject, a voiced regionsdetection (VRD) algorithm based on the frequency distribution of theaverage energy and the short-time zero-crossing rate is proposed firstly,and this VRD algorithm is not very exactly, because the threshold of thefrequency distribution of the average energy is constant, sometimesmainly rely on the short-time zero-crossing rate. Based on Teager energyoperator can efficiently detect the 'energy' of a signal, then another VRD algorithm based on TEO and lifting wavelet transform is proposed.Through a lot of experiments and simulations, results show that this VRDalgorithm can more accurately detect voiced regions and is more robust towhite Gaussian noise.(4) On the detect pitch period in voiced regions subject, in order toimprove the accuracy of pitch detection in low SNR, an algorithm basedon spatial correlation function for estimating pitch frequency only invoiced regions is proposed. Experiments show that this method iseffective. The correlation function used in lifting wavelet-based pitchdetection can sharpen and enhance sharp edges while suppressing noise.As mentioned earlier, the spatial correlation function is quite suitable forpitch detection in low SNR.