Research Of Closed-set Voiceprint Recognition System Of Text-independent

Nowadays, Biological authentication technology goes deep into people’ssocial and daily life gradually, it becomes a kind of safety verification mode byits unique convenient and economical. Voiceprint is human’s special biologicalcharacteristics, as fingerprint and face. Voiceprint recognition (namely speakerrecognition, SR) is a kind of identity authentication technology using voiceprintcharacteristic, its recognition technology is convenient and direct in comparisonwith other biological authentication technology. Voiceprint recognition isdifferent with speech recognition, it pays attention to the speaker’s voicecharacteristic in the voice signal, While, does not need to know the words’meaning of the voice. Because of that everybody’s voice characteristic isdifferent from the others (has uniqueness), not easy to forged and pretend to be,so voiceprint recognition technology is safe, accurate and reliable in the use ofidentity authentication.By recognition task, Voiceprint recognition technology is divided intospeaker identification and speaker verification; by recognition content, it isdivided into text-dependent and text-independent. This paper’s main research isclosed-set speaker identification system, aiming at putting forward new view toimprove the recognition rate on the basis of predecessors’ research.Firstly, this paper studies the research background and prospect of thespeaker identification system together with the basic structure and principle;Secondly, does a detailed study of endpoint detection, feature extraction and theGaussian mixture model (GMM). Endpoint detection module mainly puts forwarda variable step search algorithm of endpoint detection and double thresholdendpoint detection algorithm which is on the basic of spectrum distance andshort-time zero-crossing rate. Feature extraction module proposes new featureparameters that are performed on the basis of the secondary features extracted based on characteristic of MFCC and LPCC. And that is voice tone characteristicparameters. Meanwhile, does simulation experimental to compare to the variouscharacteristic parameters. Lastly, does a detailed study of the Gaussian mixturemodel (GMM), and apply it to speaker recognition model training andrecognition systems. Modifies the original Gaussian mixture model trainingalgorithm. So the model training time is greatly reduced and the efficiency of thesystem is improved. At the same time, simulates the endpoint detection algorithmand feature extraction algorithm in Chapter2and3, and the results wereanalyzed.