Design And Implementation Of Intelligent Chinese Speech Recognition System Based On FPGA

Speech recognition is one of the key technologies of human-computer interaction.The advantages of automatic speech recognition are rapid response,convenient operation,and low learning cost for users.It can be applied to C4ISR system.Recognition accuracy is a very important indicator.In order to improve accuracy,neural networks are used in speech recognition technology.However,models with high recognition rates are more complex,which poses a great challenge to the computing power of domestic processors.In order to smoothly use the speech recognition system on domestic platform,the current mainstream method is to use heterogeneous devices for computing acceleration.FPGA has low power consumption,high parallelism,and strong flexibility.It can support high-performance computing acceleration for speech recognition applications on domestic platforms.Therefore,this paper proposes and implements a speech recognition system architecture with software and hardware cooperative computing,which based on domestic FPGA.The main work is as follows.1)Optimized the architecture of the speech recognition system.On the basis of the traditional speech recognition system architecture,we have added an acoustic model based on gated convolutional neural networks.In order to transplant this system on the domestic FPGA,we only left the neural network model and vocabulary in the system architecture.It greatly reduces the cost of designing the system.And the code is easy to understand and reproduce.2)Designed the FPGA algorithm module.The algorithm module is divided into two parts:convolutional neural network operator and gated linear operator.The convolution operator is designed by modular approach.Each module is equipped with 32 computing units.According to the number of weights and step length,the number of calculation units used and the number of accumulations of each layer are configured.And then,the activation function of the gated linear operator is simplified.3)Designed the data flow control module of FPGA.Five state machines are designed in the data flow control module.The data stream and operators are scheduled and controlled through the state machines to realize the flexible and accelerated calculation of the network model.4)Implemented a speech recognition system based on domestic CPU+FPGA.A Chinese speech recognition system based on gated neural network was implemented on the domestic FT 1500A and FPGA.AISHELL data set is used to train and test the module.The experimental results show that the computing performance of the recognition system implemented on the CPU and FPGA platform is 18.1 times higher than that of the FT 1500A CPU and 5.7 times higher than that of the Intel X86 CPU.