Authentication system has been a promising field and been studied for a long time. Speech-based authentication system, part of speech processing technology, benefits a lot from the research in the field of Automatic Speech Recognition ,and is becoming some reasonable selection for information security. However, this technology still has some Achilles' Heel which focuse on how to get more identity-related and more robust information, and how to improve training method to provide models with higher generalization. This paper presents in detail the author's research work on those problems as a master candidate. Firstly, we do some deep research on Speaker Recognition (Voiceprint Recognition). especially Text-Independent Speaker Recognition. After detailedly analyzing the influence of each module of acoustic modeling, we proposed a novel method, namely Compo- nent-Group GMM, to improve the acoustic modeling, which effectively restrains the discrepancy between physical meaning and units in the elements of feature vectors. On the platform of Speaker Verification, we proposed quicker method to train anti-models and more flexible threshold setting. Hereafter, in order to find some stronger classifer, we compared GMM and Support Vector Machine to model acoustic features. We proposed embedded GMM to modify kernel functions and improve the training procedure of SVM. Secondly, we studied Verbal Information Verification which is argued that it's more robust in the comparison with voiceprint verification. We implemented Utterance Segmentation, Sub-word Hypothesis Testing and Sentence Confidence Measure. We proposed four methods to train sub-word anti-model and compared their performance in detailed testing. Thirdly, we proposed to apply SVM to integrate Speaker Verification and Verbal Information Verification, which benefis from voiceprint and user profile information and achieve satisfying performance. After that, we proposed nonstandard SVM training method to tackle the difficulty of adjusting two kinds of error-FA and FR. The new method enjoys more con- figurable authentication system.
修改评论