Research On Speaker Verification Based On Telephone Conversation

Text-Independent speaker verification is one of the most important research directions of speaker recognition and in order to contribute to the direction of research efforts and the calibration of technical capabilities of text-independent speaker recognition, National Institute of Standard and Technology (NIST) has been coordinating Speaker Recognition Evaluations since 1996. In order to explore and to seek for the suitable resolution under different conditions, it sets up different tasks and it supplies the uniform speech data, evaluation criterion and so on. As one of the tasks of NIST SRE, Summed-channel conversations speaker verification is one of the meaningful directions for research.The verification system for one speaker is introduced much detail, and then, comparing the difference between one speaker verification and conversation speaker verification This thesis make some improves from 2 points :First, comparing with the one speaker verification, Segmentation and clustering is the most important operation for conversation speaker verification , so , this thesis introduces the methods of segmentation and clustering and then proposals some methods to improve the segment results , which can get much more clean segment speech and improve the verification results.Second, the method which effectively for one speaker verification system can improve the performance of conversation speaker verification, the conception of speech quality is introduced and then , a new score method is proposed to improve the verification performance.This thesis proposes two method to improve the segment results,First, after segmentation and clustering, the purify of each segment speech is calculated, the segments with low score are discarded while the high score segments, which mean clean speech, are saved and use these speech for verification.Second, proposed a segmentation method based on fusion strategy. First, segment the speech with more than two segment methods and fusion the segment results. The same segment areas can be considered as clean speech, train the speaker models with these speech segments, and then, label the different segment areas with Viterbi algorithm.This thesis proposes two methods to deal the segment results, which can improve the system performance. First, after segment and clustering , the purify of the result is calculate, the segment with low score segments will be discard and the high score segments, which means clean speech ,will be saved ,then we use these speech for verificationSecond we can segment the speech with more then two method and then fusion the segment results. The same areas can be consider as cleaner speech ,we train the speaker models with these speech ,and then ,we label the different area with Viterbi algorithmIn order to improve the systems performance, the conception of speech quality is introduced and then, a new score method is proposed based on GMM-UBM system. by integrating auxiliary side information, the Log-likelihood Ratio is calculated dynamically . the experiments indicate the good performance of proposed method .This thesis designed a two speaker verification system, and given the performance through some different conditions, the results show the robustness and validity of the system..