The Research And Application Of Text-Independent Speaker Recognition Technology

Speaker recognition is also known as voiceprint identification, its purpose is to identify or confirm speaker identity based on the speaker’s voice characteristics. With the development of the computer and network information technology, the identity authentication’s simpleness and speed become more and more important, the speaker recognition is a biometric authentication technology, due to it’s several advantages, it has been more and more frequently applied in the fields of speaker monitor, identity verification, financial security and justice of criminal investigation and other fields, it is also the hotspot in the current voice signal identify areas. Research on the speech signal feature extraction and pattern matching problem are the key to high accuracy speaker recognition technology. The feature extraction and optimization is particularly important, feature characteristics directly influence the performance of speaker recognition system.The main research contents of the thesis are as follows:(1)Summarizing the principles of the speaker recognition technology, development and current situation, research hotspot and difficulty at present, discusses several kinds of speech features and pattern recognition methods.(2)Worked on the most currently used speech features in speaker recognition and speech recognition, did research on the identification ability and robustness and other aspects of the features, selecting proper features and fused them. This paper frame combined the SCF and MFCC features because the correlation degree between them is not high, applied score level combination between SCM feature and the former two features.(3)Researched the robustness of speaker recognition system, focused on the speaker recognition system front-end input speech forehead processing, solved the training speech recognition mismatching problem caused by additive noise and the telephone channel effects. the speech enhancement technology and improved technology was applied to solve mismatch caused by addictive noise, select a reasonable method as the denoising method. On the telephone channel brings effects mismatching problem, used various feature domain normalization techniques to improve the robustness.(4)Built a complete robust text independent speaker recognition system based on Sphinx4front-end which was finished by Carnegie Mellon University, used the technology we researched on such as speech enhancement, feature domain normalization techniques. Back end was finished using java, it can finish the job of online and offline training, online identification, as well as the recognition rate test function. And finished system testing, verify the rationality of the design and make sure it can work very well.