Research On Text-independent Multi-speaker Verification

In recent years, in the field of biometric recognition, speaker recognition attracts more and more attention for its unique security, economy and accuracy advantages, and it gradually becomes an important way to authenticate in people’s life and work. It has broad market prospect. An important research branch of speaker recognition is speaker verification. In this paper, we mainly research speaker verification technology.Starting from speaker verification system framework, the parts of system are introduced much detail. Then, to solve the speaker verification issues under complex conditions, our research focuses on the technology of feature extraction, speaker segmentation, speaker model training and so on. The main work and innovations are as follows:1. First of all, we build a automatic speaker verification system based on GMM-UBM structure as our baseline system. We study and analysis the relevant factors that affect system performance, which are the number of Gaussian function, the length of training speech and the normalization techniques. Finally we verify the effect by experiments.2. In the feature extraction, to improve the performance of speaker verification system in the noisy environment, we propose a MFCC extraction algorithm which combines multitaper and spectral subtraction methods. Improved algorithm is based on multitaper MFCC. Simulation results show that when the test speech contains the additive noise, new algorithm could achieve better result in EER and minDCF which are the evaluation index of speaker verification system.3. In the speaker segmentation, traditional speaker segmentation algorithm based on BIC exists the following defects, a large cumulative computation and too many redundancy split points, resulting in slower operation and reducing the segmentation accuracy. Relevant paper uses divide-and-conquer algorithm to improve the BIC segmentation algorithm. Although it can greatly improve the speed of BIC segmentation algorithm, at the same time, it possibly deteriorates the accuracy. In order to improve the segmentation speed and accuracy at the same time, we first propose a three-step BIC segmentation algorithm implementation strategy, then we introduce the divide-and-conquer method to the three-step BIC segmentation algorithm. Experimental results show that improved segmentation algorithm has some improvement in segmentation speed and accuracy.4. In the speaker model training, we study i-vector speaker modeling technique and focus on the extraction process of i-vector. Then we build the speaker verification system based on i-vector which is state-of-the-art technology in this research field. Then we compare the experimental results of two systems based on i-vector and GMM-UBM.