The Design Of Sound Source Target Recognition Accelerator Based On Deep Learning

It's an important way for humans to perceive the surrounding environment that is hearing.We can obtain a lot of information by listening to sounds.Voice recognition is currently used in many fields and has great value in production and life.Among them,voice recognition technology is relatively mature,and environmental sound is far inferior to voice recognition due to its complexity,maturity and application.In the early stage,voice recognition was realized by extracting the traditional acoustic features of audio,and then using statistical analysis to classify it.In recent years,Deep Learning has developed rapidly and has been applied to many fields including voice recognition.The application of deep learning in the field of voice recognition has greatly improved the accuracy of environmental voice recognition.Artificial neural network is the foundation of Deep Learning.As the application scenarios become more and more complex,neural network algorithms have become more complex and the scale continues to increase.Voice recognition is an application-based technology.Only with high accuracy and fast recognition speed can it be practically applied.At present,the computing speed,storage resources,and bandwidth of general-purpose processors relying on instruction streams are difficult to meet the computing needs of increasingly deeper neural networks.Therefore,everyone has turned to Application Specific Integrated Circuit(ASIC)and Field Programmable Gate Array(FPGA)and other hardware platforms.On the hardware platform,data calculation does not rely on instructions,but moves in a pipeline manner,so the calculation speed is faster than the processor.For identifying ten sound source targets of environmental sound data Ubandsound8 K rapidly and accurately,this paper combines deep learning with special computing hardware,and designs a special hardware computing circuit based on deep learning.The system recognition scheme is achieved by extracting the Mel Spectrum of the audio,and then inputting it into Convolutional Neural Networks(CNN)to extract advanced features to classify.The hardware accelerator designed and implemented includes two modules,namely the Mel spectrum extraction module and the CNN module.The Mel spectrum extraction module is composed of five sub-modules: pre-emphasis module,framing and windowing module,FFT calculation module,Mel filters module and control module.The CNN module has designed a hardware computing architecture for layer-by-layer computing for the goal of forward acceleration.A shared calculation module is designed in the CNN module.All convolutional layers and fully connected layers are calculated by this module.Before calculation,the shared calculation module is configured according to the control signal of the control module.The paper analyzes the characteristics of CNN algorithms,and designs three parallel computing methods: input channel parallelism,output channel parallelism,and convolution kernel parallelism according to the characteristics of each layer,and realizes the multiplexing of data in space.Finally,the accelerator system design,synthesis,and bottom-up module simulation verification are implemented in the Xilinx Vivado design environment.For comparing the recognition speed of the accelerator,the same network model is used to identify the pictures on the CPU,GPU,and the dedicated computing circuit designed in this article.The sound source recognition accelerator designed and implemented in this paper has high parallelism,high recognition performance and recognition speed.Compared with the CPU and GPU used in the test,the recognition time is reduced by 71.9% and 38%,respectively.