| Voiceprint recognition is one of the mainstream biometric identification technologies.Its advantages such as non-contact,convenience,and low cost have led to an increasing number of applications of speaker recognition in identity verification scenarios.Currently,it is widely used in fields such as security,healthcare,finance,and human-computer interaction.Among them,speaker recognition applications that are unrelated to text are the most widely used.However,existing speaker recognition models still have problems such as insufficient feature representation ability,weak generalization ability,and insufficient robustness to noise.This paper conducts research targeting the above issues,and the research results are as follows.An approach to tackle the challenges of inadequate utilization of shallow features,insufficient feature representation,and weak generalization ability is the proposal of a feature extraction network module named SE-DR-Res2 Block.This module is based on dense structures and residual structures.Firstly,the advantages and disadvantages of dense and residual structures are introduced.Secondly,a new feature extraction module DR-Res2 Block is proposed by improving the residual structure based on the feature reuse of dense structure.By fusing different scale feature information,this module enhances the attention to low semantic features and improves the feature representation ability of the model.Then,SE-DR-Res2 Block module is formed by strengthening channel attention.This module enhances the attention to important features by assigning different weights to different features,thereby improving the generalization ability.Finally,the research results based on the Voxceleb1 and SITW datasets show that the model’s error rates are 2.24% and 3.65%,respectively,which verifies that the module has good feature representation ability and generalization ability.A novel classification loss function,DV-Softmax,has been proposed based on mining loss functions and MV-Softmax to address the problems of weak noise robustness and neglecting sample quality.Firstly,the existing classification loss functions for voiceprint recognition,such as the boundary loss function,are introduced.Secondly,mining loss functions and MV-Softmax from semantic segmentation and face recognition fields are introduced.On the basis of mining loss function for simple and hard samples,the concept of fuzzy samples is proposed.Then,a more suitable classification loss function for voiceprint recognition called Dual DV-Softmax is proposed by combining mining loss function and MV-Softmax.This loss function emphasizes the quality of different samples and the relationship between classifiers to enhance model training.Finally,the research results based on the Voxceleb1 and SITW datasets show that the model’s error rates are 2.29% and 3.70%,respectively,which verifies the noise robustness of this loss function.Finally,to evaluate the effectiveness of the combined SE-DR-Res2 Block and DV-Softmax model and further address the issue of insufficient integration between model algorithms and practical applications,the SMART EAR system was built based on the combination of SE-DRRes2 Block and DV-Softmax models.Firstly,voiceprint recognition experiments were conducted on the SE-DR-Res2 Block and DV-Softmax models in different combinations based on the ECAPA-TDNN model.The research results based on the Voxceleb1 and SITW datasets show that the equal error rates of the combination forms are 2.18% and 3.59% respectively.Then,the performance of the combined model was verified on voiceprint data collected in realworld scenarios.In the voiceprint verification experiment,the accuracy rates under scenarios with light noise,moderate noise,and heavy noise were 98.9%,96.3%,and 92.1% respectively,and the equal error rates in the voiceprint recognition experiment were 1.98%,2.36%,and 3.21%respectively.Finally,based on the combined model,API interfaces for voiceprint verification,voiceprint recognition,and result retrieval were developed,as well as their corresponding GUI interfaces.The experimental results show that the combination form is effective and the SMART EAR system has good real-time performance,efficiency,and accuracy in processing voiceprint data,demonstrating the feasibility and practicality of model algorithms and application systems. |
