So you have sensors all over the fingers, use that information to get the current gesture vector, then find the basis vector that makes up the majority of the c

So you have sensors all over the fingers, use that information to get the current gesture vector, then find the basis vector that makes up the majority of the current gesture, look up the word in a Map(gesture -> word), then pass the word to a TTS engine. That'd be a toy model. What they actually do is probably more nuanced and precise.