Acoustic Model Of Speech Recognition Based On Lightweight Neural Network And Its Application In Robot

Currently,speech recognition has been widely researched and applied,and coupling neural networks to acoustic models can achieve better system performance.However,in practical applications,the computing power and storage capacity that the hardware can provide are limited.Therefore,while ensuring the performance of the speech recognition system,how to adjust to adapt to small storage,hardware equipment with low computing power has become a problem that needs to be studied and solved.And,this paper studies two lightweight neural networks,couples them to acoustic models,and builds and optimizes a speech recognition system for adaptive robot hardware devices.First,the sparse neural network is applied to the speech recognition acoustic model.A step-cut method is proposed.This method determines the size of the neurons in the neural network by specific evaluation indicators,and obtains a lightweight neural network by pruning and retraining.The method was verified on the Wall Street Journal speech data set.With only a slight decrease in system performance,the acoustic model size was reduced from 77 M to 21 M,greatly reducing the storage and computing power required by the speech recognition system.Secondly,the structural growth type lightweight neural network is studied.Contrary to the idea of sparse neural networks,structural growth neural networks control the size and performance of neural networks by gradually increasing the structure of small neural networks.Aiming at the actual application scenarios of speech recognition,based on the theory of Ada Boost algorithm,a new Con Boost structure-increasing lightweight neural network is proposed.The method was verified on the Wall Street Journal's voice data set.The results show that the Con Boost algorithm enables the neural network to reduce the number of connections,but instead gets a 8.2% performance improvement.Then,a speech recognition system suitable for embedded devices on the robot is designed and built.On the basis of the above research,for the hardware equipment provided by the robot,using the proposed Con Boost structural growth lightweight neural network,and based on the Kaldi speech recognition toolkit,the speech recognition system was designed and optimized,including the feature processing module Acoustic model,dictionary and language model combined decoder.The size of the acoustic model is only9.9M,which meets the hardware requirements.Finally,the research work and results of the full text are summarized,and several points worthy of further research are put forward.