Keyword Spotting For Lightweight Cyberphysical Chips

With the advent of the Internet of Things,smart-thing products have gradually entered people's lives,and keyword spotting technology can play an important role in the end-point products.Because people can completely interact with the end-point device in natural form,but due to the limited computing power and memory of the end-point device,and the speech recognition often has a large model and high data computation load,the voice keyword recognition scheme is very difficult to apply to the end devices.The purpose of this thesis is to design a keyword spotting solution for the end-point chip.In order to solve the problems of large model and high computation load of keyword spotting algorithms.In this work,methods such as model algorithm optimization,instruction optimization,model compression,and algorithm mapping are proposed.Finally,the keyword spotting solution meets the requirements of high precision,low power consumption,small memory footprint,and low calculation amount of the end-point device.The main work of this thesis is as follows:(1)In order to reduce noise interference in complex environments,the preprocessing stage of the keyword spotting algorithm in this thesis uses sound source localization and sound source separation algorithms to achieve noise reduction,and then uses Mel-Frequency Cepstral Coefficients(MFCC)extracts the features of the speech signal,and finally the preprocessing algorithm is transplanted into the RI5CY processor,and the preprocessing algorithm is optimized according to the extended instruction set of RISC-V.(2)In order to reduce the cost of keyword spotting algorithm and improve the accuracy of the algorithm,the speech keyword network in this thesis is an end-to-end network structure,which is composed of acoustic model and attention mechanism.After testing the network model,the model is quantified and compressed,and then the algorithm is mapped and optimized for the dedicated artificial intelligence processor.(3)In order to reduce the amount of calculation,reduce power consumption and reduce memory consumption,this thesis uses a sliding window mechanism in the inference process of the speech keyword recognition algorithm,and only one frame of data is calculated at each time.Then use multiply-add operation instead of integral operation to find the probability of random variable.The keyword detection system achieves 1.05%false rejection rate(FRR)at 1.0 false alarm(FA)per hour in this thesis.By comparing this accuracy with known solutions,it can meet the requirements of product level.In terms of the amount of calculation of the model and the speed of data operation,this thesis compresses the calculation of the model by 16 times through the quantization method,and the speed of the keyword recognition algorithm in this thesis is basically the same as the GPU operation speed through the mapping method of the dedicated processor.