Design of an Efficient Convolutional Neural Network Based on Temporal Convolution for Wake-up Spotting

In a speech recognition system that must respond to human speech in real time and process it’s the response of the wake-up work spotting to detect predefined “wake-up word” must be immediate and accurate. Previous methods for the wake-up word spotting require a lot of computation amount and a great memory consumption for high accuracy, thereby it is not very suitable to use those methods for embedded systems of limited performance. In this paper, we design a convolutional neural network (CNN) model that applies the temporal convolution layer to the structure of residual neural network (RNN), and propose a method to increase the speech and accuracy of the wake-up word spotting using that model. We compare the results by the proposed method with those by the previous methods using the Google speech command database and several representative accuracy evaluation indices of the wake-up word spotting neural network. Moreover, the efficiency of the proposed temporal convolution method is verified through comparative experiments with neural network models of identical neural network structure. The experimental results show that the proposed CNN method using the temporal convolution layer is more efficient than the previous methods for the wake-up spotting in terms of the accuracy and run-time.