LipNet: End-to-End Sentence-level Lipreading

Lipreading is the task of decoding text from the movement of a speaker's mouth. Traditional approaches separated the problem into two stages: designing or learning visual features, and prediction. More recent deep lipreading approaches are end-to-end trainable (Wand et al., 2016; Chung & Zisserman, 2016a). However, existing work on models trained end-to-end perform only word classification, rather

[suzukiMY]

http://www.oxml.co.uk/publications/2016-Assael_Shillingford_LipNet.pdf『オックスフォード大学は、読唇術専用ソフトウェアによって93.4％の読み取り精度を達成』LipNet: How easy do you think lipreading is? https://www.youtube.com/watch?v=fa5QGremQf8