The video data is probably good for training models, including text models.