Research On Speaker Recognition Method Based On AP And GMM

Speaker recognition is also called voiceprint recognition, is an important direction of the biometric identification. In recent years, with the rapid development of biotechnology and information technology, speech recognition technology has become more mature and gradually gained popularity, the voice communication with the computer has become a reality. Therefore, research on theory and algorithms for speaker recognition has been put forward and improved. Among them, the Gauss Mixture Model (GMM) because of its good performance, simple and complexity, in the field of speaker recognition is the most widely used. The classical GMM model was established for each speaker, the mixing order K associated with the specific application, generally determined by experiment, so it has great randomness. In speaker recognition process, according to the great randomness of the determining the order number of the classical GMM model, affinity propagation clustering (AP clustering) was recommended to get the order of GMM automatically, and then realize the speaker recognition method. This way can not only get the mixed order of GMM which means avoid those human works, but also significantly improve the efficiency and accuracy of speaker recognition. The main content of this paper as follows:Firstly, this paper introduced the background of speaker recognition, as well as the research significance. It also analyses the present research situation, development trend and existing problems of speaker recognition.Secondly, this paper describes in detail the process of preprocessing and feature extraction of speech signal. Among them, the pretreatment process of speech signal includes voice sampling, digital, pre emphasis, adding window and endpoint detection. Then this paper introduces the classification of speech signal feature parameters, and analysis of the linear prediction coefficient (LPC), linear prediction cepstrum coefficient (LPCC), the Mel frequency cepstral coefficients (MFCC) of advantages and disadvantages of the three kinds of feature parameters.Thirdly, in this paper, in-depth study of the process of establishing the model speaker, elaborated the model with vector quantization (VQ), hidden Markov model (HMM) and the Gauss mixture model (GMM) the basic principle of commonly used modeling methods.Fourthly, this paper gives the design flow chart of the speaker recognition system. Then, introduces AP (Affinity Propagation, AP) basic principle of clustering algorithm, and gives the concrete realization process of the automatic acquisition of GMM mixing order by using the AP clustering algorithm. Firstly, the speech feature parameters were extracted by combining the Mel frequency cepstrum coefficient (MFCC) with the differential cepstrum. Secondly, the affinity propagation clustering (AP clustering) method was used as the clustering of the speech feature parameters, and then obtained the best steps of GMM automatically. On this basis, GMM model was trained.Fifthly, this paper experiments on the model and analyzes the experimental results.In this paper, the trained GMM was used for recognizing experiment of speakers on Timit standard speech library and homemade network volunteers’ speech library. According to the experimental results, verify the effectiveness of the AP clustering and the rationality of the mixed order of GMM. The experimental results show that, the methods of GMM classic order number K was 32 and by AP clustering algorithm to obtain the order of K, the recognition time were 0.06s and 0.09s,the recognition accuracy were 90.4% and 97.6% respectively. Therefore, using the AP clustering algorithm method to automatically obtain the mixing order K of GMM, it has better recognition effect than by experiment and experience value selection the mixed order K, and the recognition time roughly the same. In other words, recommend AP clustering algorithm to get the order of GMM automatically can improve the accuracy and efficiency of speaker recognition significantly.Finally, this paper summarizes those works done in experiment. And give some advices to the further study.