Research On Improvement Of Speaker Recognition Algorithms Based On SONAR Platform

The thesis presents and implements a novel module-based Speaker recOgnitioN softwAre platfoRm (SONAR). SONAR offers speaker voice database collection, performance test and the algorithmic comparision of automatic speaker recognition. It'll apply widelier and widelier.SONAR both supports Automatic Speaker Recognition and Automatic Speech Recognition. It provides many Automatic Speaker Recognition methods and arithmetics, including speech collection, predealing, feature extraction, speaker modeling and gainning the recognition result. Furthermore, the thesis brings forward and fulfilles some new methods and arithmetics in the design and implement of the SONAR.1. Speech Filtered Recognition, a novel Automatic Speaker Recognition, and discuss its goodness and badness. Applying for Voice Filtered Recognition, we can limite the voice by verifying the content of the voice. Thus filteres the influence of the complex background and enhance performance. The Voice Filtered Recognition is an important supplement for the original methods: Speech-depended Speaker Recognition and Speech-independed Speaker Recognition.2. Further Feature ExtractionThe traditional feature extraction seems good to recognize speaker by their voice in small case. Nevertheless, the performance of the recognizer descend, owing to feature overlay limit, correlative to the increasing of the amount of user. Now we find a novel further feature extraction method, which integrate with weight, differential, combination and selection, to mine those hide personal characteristic in voice. Besides, its performance is better then the original, which is resulted from the 138-person YOHO database.3. Model Score Normalization and Gradually ScoringThe Author presents an original Model Score Nomalize (MSN) to resolve the dependent of the model and a novel Gradually Scoring (GS) to decrease the complex of the computation of the Equal Error Rate. Comparing to the normal speaker verification, MSN (including the Global MSN, the Anti-global MSN and the Trained MSN) enhance the performance. And the paper shows that the GS method quickensthe speed of gainning EER.Besides, the Author develops two application systems: the Speaker Identification System in Connected Environment and the Interactive Phone Verification System. The former is developed for many-connected methods in connected environment.The voice of in connected environment such as PC,PDA,phone,mobile and etc., can be connected to the SONAR by various pathes (sound card,transfer and voice modem,etc.).And the later support far-a-way interactive verification by phone, meanwhile it boost up human-machine capacity by combining Speech Recognizition and Text-to-Speech technologies.