Live captions
Captions as people speak, recognised on our own servers
Switched on when somebody asks for them, off the rest of the time, and the audio does not leave our machines to produce them.
What it is
Live captions under the meeting. They run when a host turns them on and stop when the meeting ends; there is no always-listening state.
The recognition runs on the same hardware as the meeting. The audio is not sent to a third party to be transcribed, which is a question procurement asks and most answers to it are "it goes to a provider".
What it costs
A flat rate per meeting, independent of how much anybody says. Measured at around a tenth of a processor core for one speaker and about one core for twelve.
What this does not do
We have never measured accuracy against a real microphone in a real room full of people. The published figures for this kind of model come from clean recordings. It is good; we will not put a percentage on it until we have measured one.