The Research Of Chinese Dialect Identification Based On Speech Feature Analysis

Dialect identification is the technology of identifying the dialect from a short utterance of the unknown speech signal which is known in some regional dialects. And it has great significance in the criminal public security investigation and speech recognition technology, and has attracted more and more attention of the researchers in related fields. The system performance of dialect identification depends on feature extraction of the speech signal, and the reasonable selection of characteristic parameters can greatly improve the recognition rate on dialect identification system. Therefore, this paper will research around the speech feature of Chinese dialect, and the main research results are listed as follows:1. A dialect identification method based on the combination of Mel frequency ceptrum coefficients and shifted delta cepstra coefficients is proposed. Firstly, in order to make the shifted delta cepstra coefficients achieve optimal performance, the optimal parameters of shifted delta cepstra coefficients on eight dialects such as Mandarin, Shanghai dialect, Cantonese, Minnan dialect, Shanxi dialect, Sichuan dialcet, Dongbei dialect and Changsha dialcet is researched. Then the combination of MFCC and SDC are employed as the feature vector with SVM for the dialect identification. The simulation results show that the dialect recognition rate with the combination of MFCC and SDC can be up to 90%.2. A dialect identification method based on S-transfrom and singular value decomposition is proposed. Short-time Fourier transform, wavelet transform and S-transform of the speech signal are researched, and then the speech signal feature is extracted using singular value decomposition from the S-transform time-frequency distribution. S-transform as the combination of the short-time Fourier transform and wavelet transform can analyze the time-frequency distribution of the speech signal, and the simulation experiment results shows that the S-transform time frequency distribution has higher resolution than short time Fourier transform and wavelet transform. Because the time-frequency distribution has high dimensions and concludes some useless information, the singular value decomposition method is employedto reduce the dimension, and identify the unknown dialect using support vector machines. The experiment simulation results show that the average rate of dialect identification based on S-transform and singular value decomposition can reach 85%.3. A dialect identification method based on linear discriminant analysis and the codebook design with GA-LBG is adopted. High dimensional speech signal means high computational complexity, and contains redundant information, which is not good for the identification, so LDA is used to reduce dimensionality. Firstly, LDA is employed to extract the feature from eight dialects such as Mandarin, Shanghai dialect, Cantonese, Minnan dialect, Shanxi dialect, Sichuan dialcet, Dongbei dialect and Changsha dialcet. Then the combination of genetic algorithm and LBG algorithm is used to design the codebook. Finally the dialect can be recognized by distortion measure.