The Design Of Speech Recognition System Based On Reconfigurable Soc Chip

In recent years, speech recognition systems on embedded systems have been widely applied to the field of intelligent home, industrial control, mobile terminals, and it is changing people’s lives. Since verbal communication is the most natural form of communication between people, human-computer interaction based embedded speech recognition systems increasingly become a hot research. However, the current automatic speech recognition system either cannot perform other tasks with high CPU utilization; or is difficult to use in embedded systems with a large volume; or is too high dependency of network which can only identify of a limited vocabulary without the network. To solve these problems, the system structure of the embedded speech recognition needs to conduct in-depth research.This thesis presents an on-chip reconfigurable speech recognition system, to a certain extent, this structure effectively alleviate these contradictions. The main work done as follows:Firstly, we study the signal processing of the speech signal. From the perspective of the signal, the thesis discusses the technologies used in the recognition process. It contains pre-emphasis, the endpoint detection, feature extraction, and other key technologies. Secondly, this thesis introduces the basic principles of hidden Markov model(HMM) and the basic principles of Gaussian mixture model(GMM). Through the exposition about the three questions of hidden Markov model, especially discussed that using the Gaussian mixture model to represent the B parameter of the hidden Markov model in detail, it has solved the principle problems of training and recognition in the speech recognition system. Thirdly, this thesis uses ZYNQ7000 as SOC design platform to build an embedded non-specific person isolated word speech recognition system. On the one hand, the PC-based training software designed before is improved to system verification platform using GMM-HMM model to provide recognition templates for the recognition system. This includes the study and implementation of training and recognition algorithms. Also includes converting data into a format which is easy to hardware testing. On the other hand, the recognition algorithm is transplanted to ZYNQ7000 platform to realize the construction of on-chip speech recognition system. This thesis includes an assessment of the recognition process, the completion of the recognition system hardware and software division, and the completion of speech recognition key algorithm’s improvements suitable for hardware features. It also includes hardware reconstruction of key computing unit, implementing digital signal processing algorithms through hardware logic.In this thesis, the main research is the reconstruction of MFCC calculation unit. Finally, through the system of recognition rate and real-time testing, we present the advantages of voice recognition system implement on the reconfigurable SOC system and the works to do in the future.