Research On Speaker Recognition Algorithm Based On Channel Compensation

As an important biometric technology,speaker recognition technology is widely used in identity authentication,information security courtroom,medical,public security and other fields.At present,the speaker recognition system using high SNR speech has reached a high recognition accuracy.However,when there is speaking channel mismatch problem,the recognition accuracy will drop dramatically.Therefore,the channel mismatch problem of the speaker recognition system is the main factor affecting the recognition performance of the speaker recognition system in real application.Targeting at this issue,this paper proposes an effective channel compensation algorithm based on the existing theory for the channel mismatch problem in speaker recognition..This paper proposes an Eigenvoice adaptive(ENV)algorithm based on noise estimation for text-independent speaker recognition systems using small data sets.The ENV algorithm uses VTS to estimate the noise characteristics in the off-line phase and combines the interference spatial projection matrix to construct a pure eigenvoice space.In the test phase,the maximum likelihood estimation algorithm is used to calculate the speaker factor that characterizes the speaker's noise-free features,and the system score is calculated using the discrete cosine distance scoring algorithm.Aiming at the difference of recognition results of i-vector and DNN/i-vector based algorithms in the score domain,this paper proposes a channel compensation(DIV)algorithm based on score-regulation weighting.The DIV algorithm effectively improves the channel compensation capability of I-vector and DNN/i-vector algorithms.In the off-line phase,the DIV algorithm calculates the degree of recognition based on the score data in the i-vector and DNN/i-vector algorithms,and then calculates the weight coefficient of the registered speaker's speech in the reference training set.In the testing phase,the weight coefficients are used to rearrange the scores of the i-vector and DNN/i-vector algorithms to calculate the final system score.In the experimental verification of this paper,we use the MATLAB to verify the ENV algorithm.Experimental results show that compared with the traditional intrinsic tone adaptive algorithm,the ENV algorithm's equivalent error rate results in the two datasets decreased by 4.4% and 1.7% respectively.The DIV algorithm is verified using the voiceprint identification tool in the KALDI voice framework.The experimental results show that for different textual and text-independent speech data sets,the performance of the system is improved by 5% to 30% compared to the performance of the system using i-vector and DNN/i-vector alone.