Exploring Semi-Supervised, Subcategory Classification and Subwords Alignment for Visual Wake Word Spotting

In this paper, we describe our approaches in Chat-scenario Chinese Lipreading (ChatCLR) Challenge task 1, which mainly explores semi-supervised, subcategory classification and subword alignment strategies for visual wake word spotting (WWS). Firstly, we add an output node to train chal-lenging samples in the negative, with a focus on confusing samples. Secondly, we employ data augmentation techniques to acquire semi-supervised data and merge it with supervised data for joint semi-supervised and supervised training. Finally, leveraging fine-grained modeling by subword alignment, we further enhance the alignment of these subwords to achieve a more precise detection. We achieved a 0.169 Score on the evaluation set, which ranks second in the ChatCLR Challenge.