Research On Text-Independent Speaker Verification System Based On Vector Quantization

With the fast development of information technology, more andmore place needs to check people's identity reliably. The traditionalidentity authentication in the form of password already exposes a lot ofdrawbacks. For insuring information security, the technology of usingpeople's peculiar biological feature as the authentication means developgradually. Speaker recognition technology belongs to biometricsrecognition technologies, which is the technology of automatic identityspeaker's authentication, according to obtain the feature parameter whichreflecting mental and physiology in voice form. Voiceprint recognitionbecome more important and popular way of ensuring security, because ofits convenience, accurate, economical and well expansibility. This thesisresearches text-independent speaker verification based on vectorquantization.Firstly, this thesis introduces the foundation acoustics knowledge ofspeaker recognition: the mathematic model of speech signal and thecharacteristic parameters of speech both in time-region andfrequency-region. And then research text-independent speakerverification based on VQ, present that union the VQ code book andaverage quantization deviation to denote the model of speaker and changethe approach to calculate the distance to the model accordingly. It isdifficult to obtain the threshold in speaker verification system, so thisthesis presents a new method to calculate the threshold of SV systembased on VQ. According to this algorithm can obtain different threshold ifthe user of system are different, and it is no need to set the thresholdbefore use. It is more robust than the custom threshold which is obtainedby experience. This developed speak verification system is simulated, andhow the common characteristic parameters affects the performance isresearched under Matlab. According to the simulation, a test platform ofSV system is built to research the performance of SV system when embedVQ or other recognition methods. Kernel algorithm of SV system isaccomplished by C language, at last. A phonemic lib, which is text-independent and has 27 persons, isrecorded in laboratory. It shows theoretically and practically that,weighted MFCC has improved recognition rate and less computation, andthe developed SV system is more robust and the false recognition rate islowed remarkably. The developed SV system has low false rate, lesscomputation and is real-time, so it can be used in identity verificationfield. But if used in special field, the performance of SV system should beimproved further.