Speaker marking method and system based on density peak value clustering and variational Bayes

Abstract

The invention provides a speaker marking method and system based on density peak value clustering and variation Bayes, and belongs to the field of voiceprint identification and mode identification. The method includes: firstly establishing a training voice database, and obtaining a general background model and a subspace model; then obtaining an i-vector factor of each fragment of to-be-detected voice data through an i-vector factor extraction method; and obtaining the number of speakers of the to-be-detected voice data and the prior probabilities of the speakers at each moment by employing a density peak value clustering algorithm, performing iterative estimation on the posterior probability of each speaker corresponding to each fragment by employing variational Bayes, and obtaining a speaker marking result. According to the method and system, problems of uncertainty of estimation of initial values of the number of the speakers and the prior probability of each speaker at each moment and large deviation generated by easy influence of the initial values on the speaker marking performance in the prior art are solved, and the accuracy, the stability and the flexibility of speaker marking are improved.

Description

technical field [0001] The invention relates to the fields of voiceprint recognition and pattern recognition, in particular to a speaker marking method and system based on density peak clustering and variational Bayesian. Background technique [0002] The significance of speaker marking technology is that when it is applied to teleconferences and international conferences, it can be saved as meeting records. At the same time, accurate identification of speakers will naturally help subsequent speech processing and semantic recognition. In addition, in the field of monitoring, speaker marking can record the voice and language of the monitored object, and when applied to the field of public security or military, it contributes to the protection of public security and even national security. [0003] Speaker marking addresses the question of who spoke when. Speaker labeling starts with extracting Mel cepstral features for speech. The Mel cepstrum feature takes into account the...

AI Technical Summary

Problems solved by technology

The invention solves the uncertainty of the initial value estimation of the number of speakers and the prior probability of the speaker at each moment in the prior art, and the speaker marking performance is easily affected by the initial value and produces large deviations; the enhancement The accuracy, stability and flexibility of speaker labeling

Images