Broadcast Attention Learning for Real Telephone Speech Keyword Spotting

Abstract With the development of mobile smart devices, keyword spotting plays an important role in interaction between machines and users. However, low storage energy consumption devices limit accuracies tasks. Therefore, how to achieve a balance high accuracy is research hotspot for system. Convolutional neural networks have been widely adopted recent systems due their superior accuracies, success transformer architecture many areas demonstrates effectiveness self-attention. In this paper, we combine self-attention convolutional networks, propose broadcast attention learning network (BA-net), using small number parameters while achieving 97.18% 78.44% respectively on Google speech command dataset real telephone dataset.