Research On Voiceprint Recognition System Based On Gaussian Mixture Model

Voiceprint recognition technology is a biometric-based recognition method.Each person's speech mode is a special feature that extracts the speaker's effective speech information from the speaker's personality speech signal.The characteristics of soundprint recognition are economical and practical,and the equipment is simple and easy to operate.It has far-reaching implications in all areas of life.The main work of this thesis is to study the voiceprint recognition technology based on Gaussian mixture model,and improve the extraction and recognition of feature parameters.After calculating the difficulty,the comparison of the recognition rates of different methods of feature extraction is carried out.Finally,the experimental results show that the improved weighted feature parameter extraction method can improve the recognition rate of the recog-nition system.According to several different stages of the voiceprint recognition system.The process of feature extraction and the construction of environmental models are studied in detail.At this stage,most of the voiceprint recognition systems are applied to the Mel frequency cepstrum coefficient based Gaussian mixture model,because the MFCC contains time-varying information of the speech signal compared to other feature extraction methods.Gaussian mixture model is one of the most commonly used modeling methods,because the GMM model has low difficulty coefficient,can be independent of text,and has stable performance.In the text,some parts of the voiceprint feature will be lost in the process of extracting the MFCC feature parameters.According to this phenomenon,a weighted MFCC is designed as MFCCw in this paper.In the experimental stage,the four processes of sampling,quantizing,pre-emphasis and windowing endpoint detection are the pre-processing stage of speech.Then,the Linear predictive cepstrum coefficient characteristic coefficient,MFCC characteristic coefficient and MFCCw characteristics of the speech are analyzed during the experiment.The coefficients were extracted and the feature coefficients extracted by the three methods were compared with a three-dimensional picture.It was found that the weighted MFCC coefficients were the most stable among the three feature coefficients,which is also the subsequent voiceprint recognition system.The recognition rate has laid a certain foundation.In the final stage of the experiment,a voiceprint recognition system based on GMM model is built.The extracted three feature coefficients are used to compare the recognition rates.The experimental results show that under the same voice duration,the weighted MFCCw method can improve the recognition rate of the system and also reduce the amount of calculation of the system to some extent.In the voiceprint recognition system based on Gaussian mixture model,the experimental data sample size is compared.It is found that the larger the training set sample in the voiceprint recognition system has a certain influence on the recognition rate,and the threshold value in the recognition process.The influence of the selection on the recognition rate is also large,and the selection of the ideal threshold can obtain a higher recognition rate.