Research On Features Extraction And Recognition Of Noisy Voiceprint Based On Fisher Fusion Algorithm

Voiceprint recognition is a biometric technology that identifies individuals by analyzing their vocal characteristics.It has the advantages of being secure,efficient,and low-cost,and has broad application prospects in criminal investigation,identity authentication,and smart homes.However,the presence of various environmental noises in daily life severely interferes with the accuracy of voiceprint recognition.Therefore,this paper optimizes the voiceprint feature extraction algorithm and improves the ECAPA-TDNN voiceprint recognition model to enhance the accuracy of voiceprint recognition in noisy environments.The main work is as follows:(1)Gaussian-Mel frequency cepstral feature extraction.In the Mel frequency cepstral feature extraction process,the Mel filter does not smooth the signal spectrum processing enough,resulting in the loss of some frequency interval features and weakening the model training effect.Therefore,this paper proposes a Gaussian-Mel frequency cepstral feature extraction algorithm,which replaces the Mel filter with a nonlinearly attenuating Gaussian filter to reduce feature errors.Experiments show that Gaussian-Mel frequency cepstral features have better recognition performance than Mel frequency cepstral features.(2)Voiceprint feature fusion based on Fisher criterion.In order to improve the voiceprint recognition performance in noisy environments,this paper proposes a voiceprint feature fusion method based on the Fisher criterion.The Gaussian-Mel frequency cepstral features and gamma tone cepstral features are fused using the Fisher criterion.The fused voiceprint features combine the high recognition performance of Gaussian-Mel frequency cepstral features and the strong noise robustness of gamma tone cepstral features,and experiments show that the fused features have better noise resistance performance.(3)Improvement of ECAPA-TDNN voiceprint recognition model.This paper addresses the problems of insufficient feature mining and inadequate multi-scale feature utilization in the ECAPA-TDNN model.The paper proposes to use a grouped convolutional residual network to extract fine-grained features and capture the dependency relationship of aggregated features in the self-attention mechanism module to improve the model’s recognition accuracy.Experimental results show that the reconstructed ECAPA-TDNN has reduced equal error rate and minimum detection cost,and the model’s performance has been effectively improved.This paper conducts research from two aspects:voiceprint feature extraction and voiceprint recognition model improvement.It improves the noise robustness of voiceprint feature parameters and enhances the performance of the voiceprint recognition model,providing a practical method for improving the accuracy of voiceprint recognition in noise environments.It has important theoretical significance for the continuous exploration of voiceprint recognition research.