Research On Background Model And Score Issues For Speaker Recognition

As a kind of biometric identification technology, speaker recognition has a unique advantage of remote control, which will be widely used with the rapid development of Internet. Many mainstream speaker recognition methods appears in recent years including GMM-SVM, JFA and i-vector, etc. These methods are based on the GMM-UBM ( Gaussian Mixture Model-Universal Background Model) framework. Therefore, the study of GMM-UBM is still meaningful. In this paper, we explored on background model construction and score issues, and proposed improvements to the GMM-UBM based speaker recognition system. The main contributions are as follows:1. We tried to build UBM using new method. We demonstrated the effectiveness of self-contained UBM, and on this basis, we proposed the concept of supporting speaker, who play a key role in constructing UBM. Dimensionality reduction by PCA, we found that the spatial dispersed speaker is helpful for finding supporting speaker set. Identify rate is about 1% higher on average than randomly selection by using this property;2. Attacks of sample copy voice for speaker recognition system limits the application and development of speaker recognition technology greatly. We propose a sample copy voice detection method based on the likelihood score changing monotonic on the order of model, utilizing the GMM model overfitting, which can be effectively used in speaker recognition system based on GMM-UBM. The correct rate of detecting sample copy voice is about 99.3%.3. GMM token is the index of gauss component who has the highest likelihood score. As a high-level features, GMM token played a great supporting role for enhance the system recognition rate. The traditional GMM token takes the highest component, and we extended it to the K highest component, which is N-Best token. In addition, we studied the different influences on the improvement of system performance on different GMM tokenizer. We validated the improved performance on GMM token ratio similarity based score corrected speaker recognition system, and achieved good results.