Unsupervised cross-domain voiceprint recognition method fusing distribution alignment and adversarial learning

Abstract

The invention discloses an unsupervised cross-domain voiceprint recognition method fusing distribution alignment and adversarial learning. The method comprises the following steps: extracting multi-dimensional acoustic features from voices of a source domain and a target domain; respectively labeling domain labels on the extracted multi-dimensional acoustic features; taking the extracted multi-dimensional acoustic features of the source domain and the target domain as training data to be sent into the network, and training to obtain classification loss of the source domain and adversarial loss of the source domain and the target domain; calculating the difference loss of the source domain and the target domain according to the domain distribution alignment loss function; calculating a loss function of the whole system according to the target function; performing gradient calculation by using stochastic gradient descent as an optimizer, performing back propagation on the gradient calculated by the loss function, and updating parameters; and through multiple iterations until convergence, completing model training. According to the method, the model can be better trained under the condition that the target domain lacks a speaker data label, so that the accuracy of cross-domain voiceprint recognition can be improved.

Description

technical field [0001] The invention relates to a cross-domain voiceprint recognition technology, in particular to an unsupervised cross-domain voiceprint recognition method that integrates distribution alignment and confrontation learning. Background technique [0002] Extracting deep voiceprint discriminative features from speech through deep learning modeling methods has become a mainstream research hotspot in this field. Deep Neural Networks (DNN) has powerful modeling capabilities and loss functions proposed for each scene, showing obvious advantages over traditional technologies. The voiceprint feature is a fixed-length vector containing the discriminative information of the voiceprint. However, this deep feature is still very sensitive to changes in the field. [0003] In practical applications, when the trained model is used in a new field, a large number of interference factors make the data distribution of the target field and the source field different, such as d...

AI Technical Summary

Problems solved by technology

However, too many network layers can easily cause gradient degradation in the training process.

Images