Using Adaptive Locally Connected Layer in Attention Based Deep Neural Network for Speech Command Recognition

Speech command recognition is an active research topic associated with the human-machine interface. Such problems can be successfully solved with attention-based deep networks. In this study, we improved one of the existing attentionbased deep network methods by using an adaptive locally connected (focused) layer. In the experiments we used Google and Kaggle datasets, which were also used in the reference. We observed that the recognition results can be improved significantly (2.6%) by the attention based deep network which uses adaptive locally connected layers.