Study On Command Word Recognition Based On Deep Learning

Speech recognition technology is more and more widely used in scenarios such as smart terminals,in-vehicle systems,and smart homes.As a branch application of keyword recognition,command word recognition can directly recognize preset command words.The command word recognition method based on deep learning has made some progress,but the existing methods still have problems.The command word recognition method based on Deep Neural Network(DNN)fails to efficiently model the local temporal and spectral correlation in the speech features,so the recognition rate is low.The command word recognition method based on Convolutional Neural Network(CNN)exploits this correlation by treating the input time-domain and spectral-domain features as an image and performing 2-D convolution operations over it,which effectively improves the recognition rate.However,it has many network parameters and a large amount of calculation.In order to solve the above problems,this paper studies the command word recognition method based on Deep learning.The main research contents are as follows:(1)A command word recognition method based on Depthwise Separable Convolution(DS)and Squeeze-and-Excitation(SE)is proposed.The Depthwise Separable Convolution effectively reduces the amount of parameters and calculations in model.The SE module learns the importance of different channels in the output feature map,calibrates the output feature map,and improves the network performance.(2)A command word recognition method based on the inverted residual with linear bottleneck structure and Squeeze-and-Excitation is proposed.When extracting features using a the inverted residual with linear bottleneck structure,the feature maps are first dimensioned in the channel dimension,and the feature maps are nonlinearly processed in higher dimensions,which can reduce the information loss brought by nonlinear operations.(3)The Ghost module in GhostNet is used to improve the network structure combining the inverted residual with linear bottleneck structure and Squeeze-and-Excitation.By using the Ghost module to replace the point convolution operation,the amount of parameters and calculations are reduced,redundant operations are reduced,and network performance is improved.