Always-on Keyword Spotting Chip Design With High Performance And Low Power

With the rapid development of artificial intelligence and the Internet of Things,voice interaction is more and more used in daily life due to its simple form and rich information.For example,Keyword Spotting System(KWS)detects the target keywords in the continuous speech,then triggers subsequent instructions.The emergence of neural network algorithms has greatly improved the accuracy of the KWS system compared to traditional models,which makes it possible to deploy the KWS system in more application scenarios.At present,the application of the KWS tends to be deployed in miniaturized smart wearable devices and large-scale intelligent sensing networks,which puts forward new requirements for the hardware implementation of the KWS system,that is,high recognition performance,low power consumption,low delay,small area,and flexible replacement of recognition tasks.However,embedded devices such as single-chip microcomputers and FPGAs are difficult to maintain due to the high power consumption during deployment,and are too large to be easily integrated.Besides,the cost of embedded devices is too high to be easily deployed on a large scale,and it is also difficult for analog chips to implement the network training algorithms with high accuracy.Thus,this thesis designs a KWS chip system by using digital circuits.The main difficulties are as follows:(i)how to deploy a neural network trained for isolated words into a KWS system for real-time continuous speech and to ensure the recognition performance;(ii)there are many types of small networks used in the KWS system,how to choose a suitable neural network algorithm model for the digital KWS chip under many constraints of software and hardware;(iii)how to maintain the high accuracy and good performance of the algorithm while realizing the digital KWS system with fixed-point with lower power consumption.To mitigate these issues,this thesis proposes the following solutions:(1)A aliasing speech feature input scheme is adopted,and a post-processing module is added.The improved KWS system is pre-deployed on the single-chip microcomputer to verify and optimize the recognition performance of the network model applied in the actual continuous speech scenarios.(2)According to the various software and hardware constraints of the KWS system in actual deployment,the comprehensive performance of the four classic neural network types is comprehensively evaluated.Finally,the GRU neural network was selected as the type of neural network deployed in the KWS system.(3)For recognition performance:the Mel Frequency Cepstral Coefficient(MFCC)feature extraction algorithm without Discrete Cosine Transform(DCT)is used to facilitate the control of the numerical range of the input features,it can improve the recognition performance while reducing the amount of calculation in the feature extraction module.For power consumption:while ensuring recognition performance,this thesis reduces the network scale,adds the fixed-point quantization,and uses a smaller clock frequency,besides,a gated clock is used to control each calculation module of the GRU neural network.Finally,FPGA functional verification and DC comprehensive performance verification show that the various performances of the digital KWS chip design all meet the expected level,which includes that 30 keywords are all awakened,and only one false awakening during the 35-minute news broadcast;the power consumption is 263W;the area is 1mm~2;and the recognition task can be switched.Compared with related research work and industrial products,this design has significant advantages.