| Keyword spotting(KWS)is a hot research direction,with a lot of application space in the fields of wearable devices,robots and smart homes.The convolutional neural network(CNN)can combine the information of the time domain and the frequency domain of the speech signal,and has good robustness to noise,so it is one of the most critical neural network algorithms in keyword spotting.However,the number of parameters and calculations of the convolutional neural network is huge,which limits the deployment of keyword spotting in portable devices.Therefore,the compression of convolutional neural networks has great practical significance.A KWS system is implemented based on CNN,and the amount of calculations and parameters of the CNN is huge.Therefore,the CNN is compressed by quantization and network pruning.By quantifying weighs,activation values and network inputs,and pruning the convolutional layer,the amount of parameters and calculations of the model can be greatly reduced.Weight is binarized to greatly compress the storage space of parameters.However,the binarization leads to the loss of network performance.The progressive quantization strategy is used to binarize the weights,which effectively reduces the performance loss of the network.The high-order residual quantization of network inputs extracts useful information from the lost information,which effectively guarantees the performance of the network.The quantization greatly reduces the amount of parameters,but does not reduce the number of calculations of the network.In order to reduce the number of calculations of the network,the convolution layer is pruned by the filter-level pruning method based on the joint evaluation strategy of the frontand-back level.Through the compression study of the KWS system,weights are compressed by 55 times,the calculation amount is reduced by about 70%,and the network accuracy rate is only reduced by 1.16%.The KWS system of the article archieves good performance. |
PDF Full Text Request