The Transplantation Research For Speaker Recognition System Of OPHONE Mobile Platform

Speaker recognition is one of the most important identifies or verifies among the methods of biometrics. By analyzing the voice from human, it could recognize the speaker. Now, the research of speaker recognition runs mostly on the X86 personal computer or the servers. As the mobile internet's development, the demand of speaker recognition on the mobile platform have expanded. The goal of this thesis is to design and implement a speaker recognition system. This system is based on large number of experiments and tests. And we use the result of experiments to improve th'e performance. The contributions of our work are the following:1 Design and implement a speaker recognition system for NIST SRE 2010We attended the NIST Speaker Recognition Evaluation 2010 and proposed a Gaussian Mixture Models-Eigen-Channel system,In the chapter 2, we introduce our system by following factor:feature Pre-processing feature extract, model train, test, score normalization. Then, we present a channel compensation algorithm, which is called Eigen-Channel, and train a channel matrix to deal with the problem of inter-session variability.2 Design and implement a speaker recognition system for mobile platformWe use existing speaker recognition system and change the parameters of the ARM system. We compare the performance between the NIST SRE system and mobile platform system. Then we reduce the order of feature and model and found the equilibrium point of recognition rate and real time. While we still improve user experience3 Propose a method which could help the system to determine a threshold quickly on the mobile platformIn order to solve the calculating shortage or computing resource insufficient, we decide to use a novel floating threshold factor and top and bottom limitation. The top limitation is computed by the training speech, the bottom limitation is generated by the imposter database. 4 Propose a speaker model update strategy and multi-speaker model determination As we known, the positions of microphone will affect the speaker model's channel and the speech's drift by time will also change the distribution of feature. Both of them will cause the worst performance. So we set a counter to record the test speech which is not the imposter. When the counter equal to a value, we use the training speech and the test speech to retrain a new speaker model. We also train difference models for difference position of microphone.