| With the rapid development of Internet technology,scientists have also started to try to process human voice characteristics through computers.Because different speakers have different vocal characteristics due to different factors such as vocal organs,age,and language habits.In recent years the field of voiceprint has developed rapidly,and the accuracy of voiceprint models has been a focus of attention.In this paper,we conduct a research on the deep learning-based vocal recognition model,based on Res Net34 residual network for feature extraction,and optimize three aspects of data enhancement,residual module and data set to achieve the purpose of improving the accuracy rate,and the main research contents are as follows:(1)Res Net34 residual network and the accuracy of the model were studied,and after 112 hours of training,56 models were finally trained,and the optimal resnet34-56 model was selected for the testing of the experiment.The effect of data enhancement on the accuracy of the voiceprint recognition model was tested using noise processing and data enhancement strategies.The model accuracy was improved by 0.8% after noise treatment,2.2% after multiple data enhancement strategies,and 2.9% when both were used simultaneously,demonstrating through experiments that data enhancement has improved the model accuracy.(2)The SE attention mechanism was studied,the SE module was added to the Res Net34 network residual module,the best trained model,the resnet34＿se-56 model,was selected,and the effect of different residual network modules on the accuracy of the voiceprint recognition model was tested.The accuracy of the model was improved by 1% after adding the SE module,and it was demonstrated experimentally that the addition of the SE attention mechanism improved the accuracy of the model.(3)The dataset was studied,based on zhvoice Chinese corpus database,adding the voice data(voice＿soft)of a class of students in the School of Software of Zhongbei University,changing the size of the training set,and testing out the effect of different training sets on the accuracy of the voice recognition model.The accuracy of resnet34-56 model improved by2.3% after the dataset was expanded.The accuracy of resnet34＿se-56 model improved by2.4% after the dataset was expanded,which proved the improvement of the model accuracy after the dataset was expanded.(4)To test the reliability and accuracy of the model,a class of 49 individuals was tested and the recognition rate reached 92.5%.This vocal recognition system fulfills the expected requirements and design well,and achieves the function of vocal pattern comparison and vocal pattern recognition of speakers. |
