The Research And Implementation Of Real-time Speaker Verification System Based On Remote Voice Communication

Speaker recognition is a kind of biological authentication technology. It can be divided into speaker identification and speaker verification, or text-independent speaker recognition and text-dependent speaker recognition. Now, text-independent speaker verification on telephone speech or voice over Internet has become one of the important fields of speaker recognition. There are some real-time speaker verification systems on service, which based on telephone or Internet network. This thesis focuses on the construction of text-independent speaker verification system based on Internet. Real-time voice active detection (VAD), UBM's training methods, establishment of speaker model and the decision strategy of ASV are mainly studied.The development and applications of real-time ASV nowadays are addressed in this thesis. The background and significance of the project are also discussed.We improve the common used double threshold VAD algorithm, which calculates the threshold of the short-time energy (E) and short-time zero cross rate (ZCR) by empirical value or the first several frames of speech data. Our new algorithm first sorts all frames' energy and ZCR, and then calculates the thresholds. Experiments show that the proposed algorithm can get good results.Under the framework of GMM-UBM, we have designed and developed a real-time ASV system based on Internet. With the help of real-time VAD, this system is evaluated based on the test data from 20 peoples. The experimental results show that the system achieved good performance in the high SNR environment.