Design And Implementation On Text-Dependent Speaker Recognition System For Short Speech

With the development of information technology, identity recognition plays a more and more important role in the field of information security insurance. Due to its stability, uniqueness and convenience, biometrics technology becomes a crucial research direction of identity recognition. As a branch of biometrics technology, speaker recognition has a good prospect in lots of fields, because of its easiness to be gathered, abundance in content and fitness to normal habits. Therefore, aiming at the application of text-dependent speaker recognition for short speech, this thesis analyzes the principle and framework of speaker recognition system in detail, then discusses the faults of the existing schemes for speech endpoint detection and speech feature matching, and gives new methods which satisfy the demands of the practical application, finally implements a system. The main contents and originalities in this thesis are listed as follows:1. Concerning with speech endpoint detection, we focus on the fact that the existing methods do not tell apart speech and low frequency noise well, and give a new speech endpoint detection scheme based on pitch and formant estimation. The experiment results show that the scheme distinguishes between speech and none-speech signals efficiently and finds out speech endpoint in Guassian noise background exactly. The performance of the scheme is superior to the other commonly used methods.2. On the side of speech feature matching, we aim at the disadvantages of existing methods, such as the invalidation of GMM under short training speech condition, and the failure of VQ to describe the space statistical character perfectly, and put forward a new fuzzy vector quantization algorithm based on orthogonal distance decomposition, as well as apply the algorithm to speaker recognition. This algorithm isn't dependent on the model-based assumption and the rule of quantization and clustering is more objective. So it acquires a better performance than GMM and VQ.3. Based on the design above, we implement the text-dependent speaker recognition system for short speech on the Visual Basic and Visual C++ platform. During the test, it is confirmed that this system achieves the design goal and gets a high performance. Hereto, we finished the research work successfully.