Study Of Pitch Detection Algorithm And The Application In Dialect Identification

As one of the important parameters of speech signal,the accurate extraction of pitch frequency is crucial for high-quality voice synthesis and analysis,speech compression coding,speech recognition as well as speaker verification.On the basis of the further study of traditional algorithm,two kinds of pitch detection algorithm with relatively high accuracy and robustness are proposed in this paper.With regard to pitch frequency,it is applied in dialect recognition system as the characteristic parameters.The research and results of the paper are as follows:(1) With the elaboration and analysis of several typical pitch detection algorithms in the world,comparison and evaluation of different algorithms are made through simulation experiments;(2) A pitch detection algorithm with a combination of the classification of un-voiced/voiced speech and the MBE self-correlation method is presented.The method uses the multi-parameter Gaussian to realize the voicing decision,and the pitch period of voiced speech is extracted by the means of MBE autocorrelation pitch detection method.The results of simulation show that the new method has better performance than that of conventional Autocorrelation algorithms,especially in the transitional parts between the surd and the sonant.It may get more accurate voice/unvoiced decisions,and improve the pitch locus.(3) A pitch detection algorithm with a combination of prediction neural model and LP-CEP is presented.The simulated pitch detection results show that the pitch extraction error of the proposed algorithm is significantly lower than that of the conventional cepstrum based algorithm both for clean speech and in low SNR of noisy speech.(4) A Chinese dialect identification system which based on a mixed SOM neural network and SVM is proposed in this paper. Hunan dialects have been selected as the research object.SOM is applied to cluster for the MFCC of various dialects,and SVM is used as the final implement of decision and identification. The results show that this system has real-time property and identification rate,especially at a low signal-to-noise ratio.