Research On Two-path BiLSTM And DCNN Model Based On Gaussian Probability Features For Speech Spoofing Detection

The current automatic speaker verfication(ASV)system is not satisfactory in response to various new voice attacks.However,many identity verification scenarios in our daily lives involve the ASV system,and the ever-increasing and changing voice spoofing attacks are the biggest threat to the ASV system.Therefore,this paper focuses on speech deception detection and constructs Gaussian probability features to improve the model's speech deception detection capabilities.The related work of this paper is summarized as follows:First of all,based on GMM independently accumulate the scores of all speech frames,while ignoring the contribution of each Gaussian component to the final score.This paper considers the score distribution of real and deceptive speech on each Gaussian component of GMM,based on linear frequency inversion.Linear Frequency Cepstral Coefficients(LFCC)and Constant Q Cepstral Coefficients(CQCC)construct Gaussian probability features to improve the effect of model speech deception detection.Secondly,this paper combines Gaussian probability features with Two-path Bidirectional Long Short-Term Memory(TBi-LSTM)to detect voice spoofing.The experimental results show that compared with the GMM model,the TBi-LSTM model has significantly improved voice deception detection capabilities.The experiment also compares the Gaussian probability features with LFCC and CQCC,and the results show that the model uses Gaussian probability features,and the voice deception detection effect is better than using LFCC and CQCC.This paper also performs score fusion on the TBi-LSTM model and the GMM model.The experimental results after the score fusion are obviously better than the results obtained before the score fusion.Finally,this paper proposes Two-path Deep Convolutional Neural Networks(TDCNN)combined with Gaussian probability features for speech deception detection.The experimental results show that the voice deception effect of the TDCNN model is better than that of the GMM and DCNN models.The experiment also compares the LFCC,CQCC and Gaussian probability features.The experimental results show that the voice deception detection effect after the model adopts the Gaussian probability feature is better than that of the LFCC and CQCC features.This paper also performs score fusion on the TDCNN model and GMM model.The experimental results show that the scoring fusion of the TDCNN model and GMM model can significantly improve the performance of speech deception detection.