Improve Hmm-based Speaker Recognition Research And Application

This thesis studies the improvement and application of H1N扞M based speaker recognition. Speaker recognition was a biometrics that recognize people via their voice, and FDvINI was the best and prevailed model in the field of speaker recogniiton.The author discuss the HiMi~vI based speaker recognition in varity aspects and fulfilled a recognition system, including speech collection, feature extraction and gain the recognition result. Some improvement have made based on the system.1.There are lots of speech parameters could model the character of the speaker, but their weights are differ. Rather the equally deal with these speech parameters, we formulate the recognition problem as a hypothesis testing problem. Implement it by training two anti-speaker models and adopt a likelihood ratio test to eliminate the parameters without the discriminating ability. Consider these two anti-speaker models based on different paradigms. they can be combined to formulate a new method, which efficiently enhance the performance.2.Using HMM, The distribution of features was formulate by density function, which didn抰 match the fact, result in recognition error. Furthmore. Conventional speaker training algo~thms are based on maximum likelihood (ML) estimation of the model distribution and the parameters of the speaker models are estimated using only the training data from the same speaker. These problems could be solved by a new formulation instead. The formulation aimed at defined a misclassification measure. minmize the misclassification error. I implement the formulation in close-set text-independent speaker identification.3.The performance of speaker rsecognition was still cant reach the real-world demand. But the speech recognition could. So. the combination the two was a feasible technique for speech base personal verification. I discussed varitv of combination idea and implement using isolated digit speech..