Research And SOPC Design Of Speaker Recognition Algorithm

Speaker recognition is a kind of biological certification technology, which automaticly recognizes speaker's identity on the basis of feature parameters extracted from speech waveform that reflect speaker's physiology and behavior characters. It is widely applied in communication, public security, finance, judiciary and other civilian centification domains due to its specialties of economy, accurary and convenience.Research work is centered on text-independent speaker recognition system. A valid and feaible miniaturization system architecture scheme is proposed based on the indepth research on pre-processing, feature extraction and recognition algorithms. Also, an open-set speaker recognition system based on SOPC is designed after rationally choosing and optimizing the whole algorithm combining with the characters of Nios II processor and FPGA. The main research contents of this thesis are as follows:1. VAD algorithm based on gassian statistic model is studied. To enhance the robustness of VAD algorithm based on statistic model, sub-band weighting algorithm based on TSNR estimation method is constructed to conquer the influence of noise and the frame-delay property of DD estimation algorithm. Experiments show that sub-band weighting methods are superior to those full-band ones of Sohn, Cho and G.729 B.2. Common feature extraction and recognition algorithms are studied. The influences to speaker recognition system from feature parameters of MFCC class and vocal source class and from two-stage recognition structure based on VQ and GMM algorithms are mainly verified and analyzed. Experiments show that: Compound feature parameters, composed of MFCC and its difference coefficients, frame logarithm energy and Renyi entropy, can efficiently trace the characters of speaker's vocal tract and souce and enable the speaker recognition system to achieve the best performance; Compound feature parameters, composed of MFCC and frame logarithm energy, make use of the least memory and recognition time and enable the system to achieve a better performance, so this kind of parameters are most suitable for embedded system implementation; This two-stage recognition algorithm reduces the calculation complexity of recognition system, while garanteeing or even excelling the accuracy of GMM recognition algorithm.3. A whole open-set speaker recognition system and a benign man-machine interchange interface are constructed based on SOPC platform, using MFCC and frame logarithm energy compound feature parameters and two-stage recognition algorithm based on VQ and GMM. And, functions, such as real-time sampling of speech, user registering with keyboard, and output displaying of system, are realized. The reliability of the speaker recognition system is verified by practical experiments.Experiments of the system show that the whole structure of speaker recognition system proposed by this thesis is feasible. Speaker recoginition system based on SOPC has unique superiority in the speed, precision and expansibility, and is an achievable solution for miniaturized speaker recognition system. And the developable space of this kind of system is broad.