Research On End-to-End Voiceprint Recognition Algorithm Based On Margin Loss

With the continuous improvement of neural network modeling capabilities and the excellent performance of deep learning on a variety of learning tasks,voiceprint recognition,as a kind of biometric technology,continues to break through technical performance.The specific features that can represent the speaker's identity are extracted from the speaker's voice,which are called voiceprint features.Voiceprint features can be widely used in the field of identity recognition.Mature applications include voiceprint locks and voiceprint recognition systems.Large-scale commercial applications are currently in progress.In recent years,the focus of voiceprint recognition has gradually shifted from traditional methods to deep learning,especially the end-to-end depth model.The existing voiceprint recognition research based on neural network has the problem of insufficient discrimination in the extracted deep voiceprint features,and it is necessary to construct more effective training criteria.When training deep neural networks for voiceprint recognition,the loss function of the model plays a key role in the convergence of the neural network.This paper analyzes and compares the recent progress of deep neural networks in the field of voiceprint recognition.Based on the general end-to-end loss function method,an end-to-end voiceprint recognition method based on margin loss is proposed.The methods based on addition cosine margin loss,addition angle margin loss and joint margin loss are proposed.By introducing an margin value,the difficulty of neural network learning voiceprint features is enhanced,forcing it to generate stronger aggregation between similar classes and different classes deeper voiceprint features with greater differences.The algorithm was verified on a large-scale English data set.By analyzing the simulation results,compared with the general end-to-end loss function,the loss function based on the additional margin makes the voiceprint recognition system have a good performance improvement in the equal error rate.The optimal equal error rate performance improvement reaches 5.3%,which fully illustrates the feasibility of this paper for the improvement method of the margin loss function.In addition,based on the voiceprint recognition algorithm based on margin loss,a voiceprint recognition system is designed and completed.By fully understanding the voiceprint recognition scene and user needs,the structure and system functions of the voiceprint recognition system are designed in detail,which successfully meets the functional requirements of the voiceprint recognition system.In the system test of the actual scene,the voiceprint recognition system designed in this paper not only meets the user's functional requirements,but also meets certain requirements in recognition accuracy and recognition speed.