Research On Voiceprint Recognition Method Based On Deep Learning

Voiceprint recognition is an important branch in the field of biometric identification.It can verify the identity of an undetermined speaker by signal preprocessing,reasonable feature extraction and an effective recognition model.It is widely applied in criminal investigation,human-computer interaction voiceprint password verification,bank voiceprint identity verification,human rehabilitation index verification and other fields.Based on the technical needs of forensic identity verification,the paper collected actual sound signal,while the signal will inevitably have strong environmental noise and stable white Gaussian equipment noise.Adaptive Wiener filtering is adopted to suppress noise effectively,and eliminate or reduce the influence to the voiceprint recognition later.After analyzed the time series and spatial features,the length of spectrum is determined to be 4 seconds to limit the spectral size as the features of voiceprint recognition.Then use GMM-UBM with opening mandarin date set Surfingtech to recognize the identity.The results show that the accuracy of voiceprint recognition is higher in the condition of determinate Wiener filter and spectrum features.On this basis,in order to reduce the redundant information in spectrum to improve the efficiency of deep learning networks in training and learning.Proposes the word-embedding spectral map dimension reduction method.And combined with the LSTM which with good timing characteristics of signal capture,a voiceprint based on word embedded LSTM using deep learning is proposed.The experimental results show the improve of the accuracy and efficiency in obviously the method above.Since the LSTM only uses the time series features of the voiceprint spectrum,the spatial characteristics are not considered.To improve the accuracy of voiceprint recognition more,making full use of the characteristics of CNN to capture the spatial information features,a idea combining LSTM and CNN is proposed.Through experimental verification,it is concluded that the spatial information of the spectrum is captured by CNN first,and the timing characteristics captured by LSTM is better for the accuracy.The results show that the word embedding dimension reduction spectrum,combined with CNN_LSTM,can get the accuracy to 97.42% in the standard speech data test set,and 95.48% in the date set with noise removed.,which was 4.25% higher than the word embedding dimension reduction LSTM voiceprint recognition method.