Microsoft Wants AI To Put Words In Your Mouth – In Realtime

Microsoft has applied for a patent on an AI system that would rewrite what a speaker says in a live video feed, and redraw their lips to match, with the changes tailored to each viewer’s profile. The filing, published Oct. 29, is an application, not a product, and the company has announced no plans to build it.

The application, US 2026/0341309 A1, is titled “Near Real-Time Artificial Intelligence (AI) Enhancement of Video or Augmented Reality (AR) Presentations of Live Events,” according to IBTimes. Microsoft Technology Licensing filed it April 29, 2025, and lists David Callaghan and Dinesh Kumar as inventors.

Per the patent tracker Patentlyze, the first claim describes an AI agent that picks “one or more spoken words to modify according to a first set of parameters,” with the choice “based on a profile of a user.” A generative AI system then swaps the words in the audio, smooths the edit and generates new video frames so the speaker’s mouth matches the new words.

Which words change depends on who is watching. Tbreak reports the profile could hold a viewer’s age, languages, region, religion, ethnicity, education, and voice or video samples. Two people watching the same broadcast could hear two different versions of what was said, as Dexerto points out.

The uses listed are mostly mundane: stripping filler words like “um,” fixing mispronunciations and stutters, replacing slang, translating speech the viewer doesn’t understand, and bleeping profanity and “offensive words.” The system is also described as pulling replacement words from an internet search, which Dexerto and Tbreak say amounts to “instant fact-checking.” That one is the least mundane. Software that decides a speaker’s statement was wrong and quietly substitutes what it thinks they should have said is a different thing from a profanity filter.

The filing calls the process “near-real-time,” which it ties to the few-second delay broadcasters already use for manual profanity filtering, with one to 10 seconds given as an example. The patent mentions lectures, award ceremonies, sporting events and live TV, and says the software could run on phones, tablets, cameras and AR glasses. Another claim would turn long recordings into shorter summary videos built partly from newly generated footage.

Only one dependent claim addresses the obvious problem. It keeps the unaltered audio and video as “ground truth,” Patentlyze says, and that is the only claim that preserves the original. Nothing in the coverage says viewers would be told a feed had been edited, or could see the original.

Status is murky. The patent examiner raised objections on July 29, 2026, and the first review rejected the claims as filed, according to Tbreak and Patentlyze. That is a routine step that usually leads to amended claims, and some outlets describe the application as further along than that. Microsoft did not respond to Futurism’s request for comment, and its shipping Teams Interpreter translates audio only and does not touch the speaker’s face.

Microsoft is not alone in the idea. Tbreak notes Sony has a similar patent for AI that replaces sensitive moments in games using parent-set rules. Critics quoted in the coverage worry that edited footage could be mistaken for an accurate record of what was said; IBTimes says that matters to journalists, broadcasters and anyone using video as evidence. Real-time face and voice swapping is already a scammer’s tool, and a patent that treats a live speaker’s words as editable per viewer shows where the large AI companies are looking.