Research On Key Techniques Of Voiceprint Recognition Under Vehicle Noise Background

In recent years,the level of voiceprint recognition technology has gradually improved with the rapid development of artificial intelligence.voiceprint recognition technology has received increasing attention.After the car has entered the households,with the demand for in-vehicle entertainment and the rapid development of electronic components,in-vehicle voice equipment has become one of the research hotspots today.Because voiceprint features are susceptible to complex environments and instability factors,voiceprint recognition systems have become an important subject in the context of vehicle noise.Voiceprint recognition refers to extracting the speaker's personal information from the speech signal to distinguish the speaker's identity.Voiceprint recognition system is mainly divided into three parts:front-end processing,feature extraction and recognition model.The quality of the speech after the front-end processing directly affects the quality of the feature parameters extracted in the next step,and the quality of the feature parameter extraction further affects the accuracy of the recognition.Therefore,the front-end processing and feature extraction are two very important parts of the voiceprint recognition system.This article first explains the basic principles of key technologies such as voiceprint recognition,voice endpoint detection,voiceprint featureextraction,and then studies some key technologies of voiceprint recognition systems under the background of vehicle noise.The main work of the thesis is:1.Based on the analysis of the advantages and disadvantages of the traditional dual-threshold endpoint detection algorithm,a combination of genetic simulated annealing(GASA),fuzzy C-mean(FCM),and Bayesian criterion(BIC)is proposed,and an optimization based on GASA is proposed.Voice endpoint detection method of FCM-BIC algorithm.In this method,short-term energy and spectral entropy are used as threshold parameters,and a genetic simulated annealing algorithm is incorporated.The obtained cluster center is assigned to the FCM-BIC to determine the threshold value of the signal characteristics.Finally,the voice endpoint is detected based on the threshold.The experimental results show that the weighted error measure of endpoint detection in this method is smaller than the traditional double-threshold method,the algorithm improvement effect is more obvious under white noise,and the endpoint detection effect is best under vehicle noise.Aiming at the traditional bottleneck feature(BN)extraction method,the anti-noise performance is not strong,and the redundant information is too high,which leads to low recognition rate.A bottleneck feature extraction method based on TCL and sparse DNN network is proposed.TCL is introduced to classify the training corpus in time structure,and anappropriate overlapping group sparse regular term is introduced to the cross-entropy-based objective function to build a sparse DNN network.Finally,experiments show that compared with traditional voiceprint features(MFCC,LPCC)and bottleneck features based on sparse deep neural networks,the equal error rate(EER)is reduced to a certain extent,which can effectively improve the accuracy of speaker recognition.3.Based on the research on endpoint detection methods and feature extraction methods in voiceprint recognition technology,these voiceprint recognition technologies are applied to speaker recognition systems under the background of vehicle noise,and the Gaussian mixing order is studied through comparative experiments.Endpoint detection and signal-to-noise ratio affect the performance of voiceprint recognition system.Finally,experiments show that the recognition rate of the GMM-UBM acoustic model is higher than that of the GMM model,and the recognition rate of the GMM-UBM model system increases with the increase of the mixing order;endpoint detection can effectively reduce the impact of noise on the recognition system.In recognition systems with short speech lengths,the speech endpoint detection method based on GASA optimized FCM-BIC improves the system recognition rate more significantly;the bottleneck characteristics based on TCL and sparse DNN networks have lower signal-to-noise ratios than traditional voiceprint parameters.The recognition rate has improved significantly under the environment.