Speaker Recognition Accelerator Design Study

Nowadays the speaker recognition system is used in a wide range, especially on the information security and the man-machine conversation. Most systems are originally designed to run on desktop computers, the recent technological and market growths of embedded systems draw our attention to enable speaker recognition system running on Application Special Integrate Circuit. Its potential applications include time-attention systems, access control devices and security modules in electronic appliances.The speaker recognition that extracts the voice features and builds the relevant models is a task to determine whether a person is who he/she claims to be based on his/her voice. It is applied for recognizing the identity of the speaker. A typical speaker recognition system is composed of four parts, which is front-end processing, feature extraction, training of speaker model and pattern matching. The paper researches an accelerator of speaker recognition system and the aim is to improve the performance of system capability. The main tasks include:First of all, the key technologies of the system and the application of vector quantization technology in the speaker recognition are analyzed. Codebook training algorithm and the parameter values are discussed in the paper. Small voice corpuses are built and the voice front-end points are detected. The whole system is simulated on the MATLAB simulation bench. This part consists of extracting Mel Frequency Cepstrum, training codebook and recognizing speaker. The result indicates the system has the high recognition rate.Secondly, an accelerator of speaker recognition is designed. The characteristic parameters that are the data used in designing and verifying the models are transformed from floated-point into fixed-point. The accelerator is composed of two models, one is codebook training model and the other is speaker judging model. Both are based on vector quantization technology.In the end, the system verification blue print is given. At the same time, all the models are compiled, simulated and synthesized. The results are analyzed and the improvements are proposed.