A newly published Microsoft patent application describes a system that would rewrite what a speaker says while they are still speaking — and redraw their mouth so the edit does not show. The patent application, filed on April 29, 2025 and published on October 29, 2026, describes what it calls near-real-time enhancement of live video, with changes landing within seconds rather than in a later edit.
The idea, first flagged by patent tracker Patentlyze and then covered by several news outlets on November 2, is not a product announcement. Microsoft has not commented on the filing, and a patent application does not confirm the company is building or planning to launch such a tool. But the document lays out, in unusual detail, how a live feed could be rewritten on the viewer's own device.
How it would work
According to the filing, cameras and microphones capture the live audio and video, and an AI agent scans the speaker's words as they arrive. It flags words to change based on a set of parameters and a profile of the viewer — things like their age, the languages they speak, or where they live. A generative AI system then replaces the chosen words in both the audio and the video, reshaping the speaker's mouth and smoothing the surrounding frames so the lips match the new speech. The filing defines near-real-time as roughly one to ten seconds, and names AR glasses, phones, tablets and cameras running on device chips as possible hosts.
Short blended transitions would hide the seams, so the viewer would hear a continuous, cleaned-up version of the talk while it is still happening. In effect, each viewer could receive a slightly different version of the same live event.
From filler words to live translation
The pitched uses range from the mundane to the ambitious. The system could strip filler words such as “um” and “you know,” erase stutters, correct mispronunciations, and remove slang, dialect terms, or words the viewer has blacklisted. It could also filter offensive or inappropriate language, and even swap in replacement words found through an internet search — with the filing floating instant fact-checking as a possibility.
Translation is the most striking use case. If a speaker uses a language outside the viewer's profile, the system could replace their words with a translated version and adjust their lip movements accordingly, so the viewer would hear and see them speaking their own language. Lectures, award shows and live sports are named as settings where the tool could apply.
The trust problem
The same capability is also the filing's biggest problem. If AI can change what someone appears to say and make their face match, viewers may struggle to tell original footage from modified footage — a risk commentators have already flagged for misinformation, since a speaker's words on an important matter could be altered in a feed that still looks completely real.
Microsoft's own document appears to nod at that danger: one dependent claim proposes recording and storing the original live video and audio as ground truth. Patent watchers also note that an examiner's first review rejected the claims as filed, which is routine and usually followed by amendments. For now, the system exists on paper only — a glimpse of how personalised live video could become, and a warning about how hard it may get to trust what you see and hear.



