Voiceprint identification method and system

Abstract

The invention provides a voiceprint identification method and system. The method comprises steps that features of voiced frames in a training corpus set are extracted through VAD voice activity detection; inter-class angle boundaries of the features of the voiced frames are expanded based on the A-softmax loss function, and the intra-class angle of the features of the voiced frames is limited to train a neural network model; deep voiceprint features of a to-be-registered target are determined according to the trained neural network model, and the to-be-registered target and the deep voiceprintfeatures are registered in a voiceprint database; the deep voiceprint features of the to-be-registered target are determined according to the trained neural network model; identification is carried out according to similarity of each deep voiceprint feature in the voiceprint database and the deep voiceprint feature of the to-be-registered target. The invention further provides a voiceprint identification system. The method is advantaged in that the A-softmax loss function is utilized to limit the intra-class angle, so obvious angle boundaries are between corresponding different classes of embedding vectors, discriminability is improved, and identification accuracy is higher.

Description

technical field [0001] The invention relates to the field of voiceprint recognition, in particular to a voiceprint recognition method and system. Background technique [0002] Voiceprint recognition refers to identifying or verifying the identity of the speaker through the voice segment provided by the speaker. According to the lexical constraints of spoken content, it can be divided into two categories, text-dependent lexical constraints and text-independent lexical constraints. [0003] For voiceprint recognition, an i-vector-based voiceprint recognition method or an end-to-end voiceprint recognition method is usually used. Among them, i-vector PLDA (Probabilistic Linear Discriminant Analysis, Probabilistic Linear Discriminant Analysis) is a relatively advanced algorithm in the field of voiceprint recognition. Under the framework of i-vector, the supervector M obtained by UBM (UniformBackground Model, universal background model) is modeled as M=m+Tw. where m is a speake...

AI Technical Summary

Problems solved by technology

[0006] In order to at least solve the poor performance of i-vector on short-term data in the prior art, the speaker representation vector trained by the end-to-end voiceprint recognition method is not distinguishable enough, and the neural network of the triplet loss function achieves full convergence speed For the slow problem, the applicant unexpectedly discovered that using the A-softmax method, which is more popular in the image field, expands the inter-class angle boundary of the speaker's voice and limits its intra-class angle

Images