Speech densely connected convolutional networks for small-footprint keyword spotting

Keyword spotting is an important task for human-computer interaction (HCI). For high privacy, the identification task needs to be performed at the edge, so the purpose of this task is to improve the accuracy as much as possible within the limited cost. This paper proposes a new keyword spotting technique by the convolutional neural network (CNN) method. It is based on the application of densely connected convolutional networks (DenseNet). To make the model smaller, we replace the normal convolution with group convolution and depthwise separable convolution. We add squeeze-and-excitation networks (SENet) to enhance the weight of important features to increase the accuracy. To investigate the effect of different convolutions on DenseNet, we built two models: SpDenseNet and SpDenseNet-L. we validated the network using the Google speech commands dataset. Our proposed network had better accuracy than the other networks even with a fewer number of parameters and floating-point operations (FLOPs). SpDenseNet could achieve an accuracy of 96.3% with 122.63 K trainable parameters and 142.7 M FLOPs. Compared to the benchmark works, only about 52% of the number of parameters and about 12% of the FLOPs are used. In addition, we varied the depth and width of the network to build a compact variant. It also outperforms other compact variants, where SpDenseNet-L-narrow could achieve an accuracy of 93.6% withiri: An On-device DNN-powere 9.27 K trainable parameters and 3.47 M FLOPs. Compared to the benchmark works, the accuracy on SpDenseNet-L-narrow is improved by 3.5%. It only uses only about 47% of the number of parameters and about 48% of the FLOPS.