Improved features using convolution-augmented transformers for keyword spotting

Transformer can effectively model long rang dependency, but suffer from uncapable to extract local feature patterns. While CNNs exploit features effectively. In this paper, we seek combine convolution and Transformers improves over using them individually, propose improved convolution-augmented transformers for keyword spotting. The are constructed with a ResNet front-end back-end in series. Using spotting task. results show that the convolution- augmented yield at least 3% improvement compared other features.