An ordinary dictation system listens to sound. A silent speech interface instead observes signals associated with forming words, such as small movements or electrical activity from muscles around the mouth and jaw.
A recognition model can learn to map those signals to words or commands. Performance depends on the sensor, the user, the vocabulary and the conditions of the test, so a successful demonstration may not establish reliable everyday dictation.
ELI5
Imagine forming the words 'next page' without speaking aloud. A sensor picks up some of the movements that go with those words.
An AI system tries to match the signals to the command and turn the page. It can make mistakes if the signals are unclear or unfamiliar.
Does this read someone's thoughts?
A system based on muscle activity needs a measurable signal associated with forming words. It does not reveal unrelated thoughts.
Is this the same as ordinary speech recognition?
Both interpret language input, but ordinary dictation uses recorded sound. Silent speech interfaces can work with other signals.

