Research On Speaker Recognition In Coded Domain

Coded domain speaker rocognition is a technique that studies speaker recognition from coded speech. The characteristic of this technique is that speech features are extracted from coding stream directly, while compressed speech in traditional speaker recognition technology is decoded first. Coded domain speaker recognition is a new field with the development of modern communication. It involves in speaker reocognition and speech coding technique and has an extensive application foreground. This dissertation mainly studies coded domain speaker recognition with three achievements as following.Firstly, this paper studies the infuence of speech coding on the performance of speaker recognition, and proposes a weighted GMM model based on coding distortion. A coder automatic matched method and score normalization technology are used in code mismatched condition. Experiments show that these motheds can improve the the accuracy and practicabilty of speaker rocognition.Secondly, three optimized feature vectors are presented in GSM FR, G.729 and G.723.1 coding stream, respectively. The process of extracting optimized feature vectors includes two steps. Basical features are extracted from the three coding stream first, then feature selection and feature combination technique is used to find the optimized feature vectors. Experiments show that optimized feature vectors are universal. Compared to the extraction from decoded speech, they can improve effectivity and efficiency of speaker recognition.Finally, a frame selection method of VoIP coded domain speaker recognition is proposed based on frame selection. In this method, silence and bad frames are detected in VoIP packet stream by analyzing the transporting protocal and the characteristic of the stream. Experiments show that the new mothed of frame selection can improve effectivity and efficiency of VoIP coded domain speaker recognition.