The Design And Implementation Of Open-Set Speaker Recognition System

Speaker recognition is a biometric authentication technology, which automatically identifies the speaker identity according to the speech waveform parameters that reflect the static and dynamic characteristics of the speaker’s voice. The basic types of speaker recognition are speaker identification and speaker verification, the former means the speaker of test data belongs to the known speaker set, you need to find out the speakers most similar with test data from the speaker set; and the latter will have to determine whether the test data actually comes from a given speaker.The speaker recognition can be divided into open or closed set identification on whether the speaker of test data is within a known set of speakers. The close-set speaker recognition can be regard as a speaker identification process, and for the open-set identification, one should not only to find the same set of test data which is most similar to the speaker, but also to confirm the identity of the test data, whether this is the speaker or a collection of strangers outside speaker. In the actual system application, we generally do not know whether the identity of an unknown voice is within the set, so the open-set recognition is recognized more widely.In this article, the open set speaker recognition is regarded as a combination of speaker verification process and identification process, and we explore the potential direction of the study that can improve the performance of these two processes, like scoring regular and training reasonable threshold to ensure the accuracy of speaker verification program, and speaker clustering based efficient search algorithm for speaker identification process to improve its performance, Based on this, the paper implements an open-set speaker recognition systems which applied the upper study of developed treatment programs and improvements to the actual system, and finally test its effectiveness through the experiment.