Speaker confirmation method adopting unsupervised clustering score regularity

Abstract

The invention discloses a speaker confirmation method adopting unsupervised clustering score regularity, which comprises the following steps of: testing a current speaker model of a registration set and a regularized data set containing L corpora to obtain L test scores; clustering the L test scores by using a K-means algorithm, and screening out a plurality of types of scores with the front meanvalues; adopting an expectation maximization algorithm to cluster a plurality of types of screened scores, introducing a Gaussian mixture model to fit score distribution for probability distribution,and after the Gaussian mixture model converges, taking a Gaussian unit mean value with the maximum mean value and a standard deviation as score regularization parameters of the current speaker model;regularizing L test scores by using score regularity parameters, and performing speaker confirmation by using the regularized scores.

Description

technical field [0001] The invention relates to the field of speech signal processing, in particular to a speaker confirmation method using unsupervised clustering score regularization. Background technique [0002] In recent years, with the rapid development of information technology, people generate more and more information, and there are more and more channels to obtain information. How to protect the privacy of personal information has become a key issue. Identity verification has become a topic worthy of attention. Speaker verification or speaker recognition, also known as voiceprint recognition, is a type of biometrics, and its purpose is to use a machine to automatically judge a segment of speech based on a given segment of speech. The process of testing whether a voice is consistent with its claimed identity. Due to the inherent differences in human physiological organs, there are also differences in acquired pronunciation methods and habits, which make each perso...

AI Technical Summary

Problems solved by technology

The aforementioned paper (D.E.Sturim and D.A.Reynolds, "Speaker adaptive cohort selection for tnorm in text-independent speaker verification" [C], / / Proc.ofICASSP, 2005:741–744) selects some of the highest-scoring impostor scores to estimate the regularization parameter , the distribution of the scores with the largest values ​​does not obey the Gaussian distribution, so there must be a deviation from the distribution of the test set; in addition, how many of the largest scores to choose is actually an empirical value, which has a great impact on the actual performance

Images