Research On Speaker Confirmation Technology Based On Pronunciation Action Parameters

Voiceprint recognition is a recognition process for target speaker based on physiological and behavioral characteristics and features of speech signal.Compared with other security authentication methods,voiceprint recognition technology has the advantages of suitable for remote speaker identity,convenient and low cost,because of that voiceprint recognition is widely used in commercial applications.In this paper,we mainly study articulatory movement features(AMFs)for short-duration text-dependent speaker verification technology.AMFs are borrowed from speech recognition area,when compared other acoustic feature coefficients with AMFs,AMFs are not influenced by channel and environmental background noises.Therefore,the AMFs can be more noise-robust.Firstly,we introduce the extraction process and regularization process of AMFs,and also introduce traditional acoustic feature extraction technology,such as Mel-Frequency Cepstral Coefficients(MFCC),Linear Predictive Coding Cepstral Coefficients(LPCC).we study the difference of acoustic signal between AMFs and MFCC,under the conditions of same dimension and same speaker.When comparing the amplitude values,we find that AMFs are superior to MFCC.Speech signal,speech feature extraction and algorithm are key factors in the study of text dependent voiceprint recognition.In this paper we build a new database.Based on this database,MFCC feature parameters of different dimensions are selected.Studying feature parameters,we choose traditional modeling methods,such as Gaussian Mixture Model-Universal Background Model(GMM_UBM),Dynamic Time Warping(DTW).we find that DTW algorithm superior to GMM_UBM algorithm.AMFs can reflect speaker identity information,due to AMFs are unaffected to channel,environment,physical condition and emotional.In order to choose the most discriminating AMFs,AMFs of different phonological organs at different reference points were selected,using DTW model to verify speaker identity.Discovered that AMFs using up lip as reference point are the best.using feature and system fusion techniques to reduce the EER of voiceprint recognition.In the end,we change the train part in speaker verification.Using Gaussian Mixture Model-Support Vector Machine(GMM-SVM)speaker verification Technology,to build speaker verification framework based on AMFs.Theexperimental results are consistent with the results of the AMF DTW-based speaker verification system,and the recognition result of the GMM_SVM system is better than DTW system.