| With the development of embedded system and speaker recognition technology, speaker recognition system based on embedded system is widely applied to the fields of medical services, bank and securities, public security and justice. In recent years with the development of mobile internet, speech may be the first available feature parameters to represent the identity characteristics of customers for the remote authentication through mobile phone, tablet computer and other mobile devices, causing an upsurge in the speaker recognition research.The system of text-independent speaker recognition system which based on Xilinx Zynq is designed, laying the foundation for the following development based on mobile devices. The main work includes:Firstly, the methods of phonetic sampling and quantizing, pre-processing and the extraction of feature parameters is introduced. The key technologies used in pre-processing is described in detail: pre-emphasis, framing and adding windows, endpoint detection. And MFCC, the method of feature parameters extraction, is introduced specifically.Secondly, the basic principles of Gaussian Mixture Model(GMM) is introduced. The Expectation Maximization(EM) algorithm is introduced specifically, which is used to train the parameters of GMM. Then the k-means algorithm is introduced which is used to initialize the training parameters. The principle of training and recognition in the speaker recognition system is introduced, and the simulation experiment using Matlab is conducted.Thirdly, the embedded text-independent speaker recognition system by taking Xilinx Zynq as the hardware platform is constructed. By introducing the hardware platform resource, decompose the software construction of the speaker recognition system into two parts: assistant software based on PC and recognition software based on embedded system, sharing files between PC and embedded system by Network File System(NFS). The assistant software based on PC is used to train the parameters of GMM and the verification of recognition algorithm. Considering the complexity of FFT calculation unit, the hardware acceleration module of FFT calculation unit is designed. The embedded environment for the hardware platform with the Real FFT hardware logic is constructed: adding a u-boot new command to parse the configuration files and modifying the u-boot source code to support booting multiple systems, modifying the device tree source code to partition flash, creating root file system and adding them into Linux kernel. Then the software interface for the hareware acceleration logic is designed.Finally, the speaker recognition system is tested, and the recognition rate and the real-time response of the system is analyzed, then the outlook for the future of research is stated. |
