Threshold setting method based on multimodality detection in speaker verification system

摘要

How to make a decision is a critical problem in speaker verification system and it directly affects the final verification results. This paper describes a new efficient threshold setting method that performs speaker verification system with I-vector technique. Unlike typical system under laboratory, it is difficult to obtain a large number of data to fully estimate the speaker recognition threshold and the speaker's score distribution is also different. Therefore, we propose a new threshold estimation algorithm based on multimodality detection and abnormal score pruning method. The experimental results show that compared with the ordinary method, the overall recognition performance of open-set speaker recognition system based on the new decision threshold is improved.