Leveraging Unimodal Self-Supervised Learning for Multimodal Audio-Visual Speech Recognition

Training Transformer-based models demands a large amount of data, whileobtaining aligned and labelled data in multimodality is rather cost-demanding,especially for audio-visual speech recognition (AVSR). Thus it makes a lot ofsense to make use of unlabelled unimodal data. On the other side, although theeffectiveness of large-scale self-supervised learning is well established inboth audio and visual modalities, how to integrate those pre-trained modelsinto a multimodal scenario remains underexplored. In this work, we successfullyleverage unimodal self-supervised learning to promote the multimodal AVSR. Inparticular, audio and visual front-ends are trained on large-scale unimodaldatasets, then we integrate components of both front-ends into a largermultimodal framework which learns to recognize parallel audio-visual data intocharacters through a combination of CTC and seq2seq decoding. We show that bothcomponents inherited from unimodal self-supervised learning cooperate well,resulting in that the multimodal framework yields competitive results throughfine-tuning. Our model is experimentally validated on both word-level andsentence-level tasks. Especially, even without an external language model, ourproposed model raises the state-of-the-art performances on the widely acceptedLip Reading Sentences 2 (LRS2) dataset by a large margin, with a relativeimprovement of 30%.