Reading speech from the movement of the lips · Ceded (since 2018)

In July 2018 a DeepMind and Oxford system reported a word error rate of 40.9 percent on 3,886 hours of unconstrained video, against 92.9 percent for professional lipreaders given the same material and extra context.

LipNet had already reached 95.2 percent sentence accuracy in November 2016, but on the GRID corpus, which uses a fixed grammar and a vocabulary of a few dozen words, so the result proved less than it appeared. The 2018 figure holds on ordinary television footage, which is why it counts as a transfer. What was ceded is the decoding of visible speech from a clear frontal view, a task at which trained people have always been poor and machines are now merely bad. Not ceded: lipreading as deaf people practise it, which runs on knowing the speaker, on the subject at hand, and on repair strategies when the reading fails, and which no word error rate has ever measured. The capability also arrives as surveillance before it arrives as help, which is a fact about the transfer rather than about the skill.