A dataset for lipreading using sequences of video frames

Lipreading is the task of decoding text from the movement of a speaker’s mouth. Based on LipNet, a model that maps a variable-length sequence of video frames to text, making use of spatiotemporal convolutions, a recurrent network, and the connectionist temporal classification loss, trained entirely end-to-end.