Research And Implementation Of Speaker Recognition Algorithm Based On Text Independent

With the rapid development of Internet technology,the network has gradually covered every corner of social life.In the Internet environment,the requirement of the identity authentication technology is getting higher and higher so that traditional identity authentication technology has been unable to meet the requirements of safe and convenient authentication.In all the methods of identity authentication,biometric identification technology is a kind of identification technology based on physiological and acquired characteristics of human being,because of its unique advantages,it has been widely used in practice.Among all the biometric identification technology,textindependent speaker verification technology is considered as one of the most pract ical identification technology,the technology uses the speech of the target speaker to confirm the identity of the speaker and is an important branch of speech recognition research.In the practical application environment,due to the influence of many factors,such as acquisition equipment,transmission lines and so on,the resulting effect ive speech data is very limited,and then it is difficult to achieve the desired effect of recognition and execution efficiency of the system.Therefore,this paper is mainly based on the textindependent voiceprint verification system research which the voice data is relatively short.In a speaker verification system,the recognition rate and computationa complexity are important indicators to measure system performance.To some extent,traditional UBM-MAP-GMM model structure solves the problem of mismatc h between test speech and training speech,and the system identification performance is also ideal.However,in practical applications,in the face of the short speech problem,the model has large amount of computation and poor robustness.So the paper will studies the voiceprint recognition algorithm mainly in order to reduce the computation complexity and improve the recognition rate,specifically in the following areas:1.In model training,the influence of the initial value of the model on the EM algorithm is analyzed.According to the defect of traditional K-means algorithm which randomly choosing initial clustering center may lead to local convergence,the initial clustering center selection algorithm based on density and distance is proposed and the feasibility of the improved K-means algorithm is proved by experiments.2.By analyzing the UBM-MAP-GMM model architecture,for its large amount of calculation,the individual voiceprint model GMM forced to follow the same model structure,the influence of partial Gauss component on recognition results,the UBM-CM-MAP-GMM model structure is proposed.The experimental results show that the improved model not only can reduce the test time but also can improve the equal error rate of the system.3.The mixing degree value of GMM is studied on UBM-CM-MAP-GMM model and the experimental data show that the best effect is when the GMM mixture is half of UBM.4.Phrases sound speaker verification system is realized based on the framework of UBM-CM-MAP-GMM model.The recognition rate of the system has been analyzed and validated by experiments.Compared to the traditional model,UBM-MAP-GMM,the improved model can reduce the computation and error rate to a certain extent.