Live broadcasts used to have one virtue: whatever the speaker said, that was what you heard. Microsoft has filed a patent for a system that would change the words while the person is still talking, and then adjust their mouth so the edit doesn't show.
What the filing describes
The patent covers AI agents that alter the audio and video of a live performance practically in real time. The audio is only half of it. The system is also meant to change the speaker's image, including lip movements and facial expression, so the picture agrees with the new version of the speech.
The editing itself works on words and phrases. The system can filter them, delete them or put different ones in their place. What gets changed would depend on predefined parameters or on user profiles. To hide the seams, the algorithm also corrects the video and sound just before and just after the altered fragment, so the transition looks and sounds natural.
When the system needs a replacement word, the description says it can find one in several ways, including an internet search.
The polite uses
The filing opens with applications that are easy to defend. One is translating a live broadcast into the language the viewer prefers. Another is cleaning up delivery by removing filler words and fixing stuttering and other speech peculiarities.
The formats listed are broad: lectures, presentations, television programs, ceremonies, shows and sports broadcasts. Anything with a person and a microphone, in other words.
The less polite uses
The list does not stop at stutters. The patent also mentions removing local expressions and colloquial turns of phrase, words that were mispronounced or used wrongly, and statements that are offensive or inappropriate.
That is quite a distance to cover in one feature list. It starts with "um" and ends with whatever a viewer's profile decides is inappropriate, and there is no visible seam in between.
The result is that a viewer could receive a personally edited version of a speech, with wording the system judges to fit that viewer's settings. Two people watching the same live event could see two slightly different speakers saying two slightly different things. Both would be, technically, live.
The safety net
The patent provides for keeping the original video and audio as a reference recording. That could help tell the unaltered broadcast from the processed one.
It does not settle the matter, though. A reference copy only helps if someone goes looking for it, and the report points out that the feature does not rule out confusion once edited recordings start circulating. A clip with a convincing face and an adjusted sentence travels faster than the archive that could contradict it.
What is still unknown
This is a patent filing, which shows what a company wants to claim and says little about what it plans to ship. The reporting available says nothing about a product, a release date or where Microsoft might use the technology. Nor does it say how the system would handle the delay that real-time editing of live speech implies.
So for now the useful reading is narrow. Microsoft has put on paper a method for changing what a person appears to say, as they say it, with their face adjusted to match. The stated reasons are translation and tidier speech. The listed reasons go further than that.
Live used to mean that what you saw was what happened. The patent proposes a revision: what you see is what your settings approved.