Microsoft Patent Application Describes AI That Could Rewrite Live Speech and Alter Lip Movements

What was actually filed​

The document is U.S. application 2026/0341309 A1, from Microsoft Technology Licensing, LLC. It is titled "Near Real-Time Artificial Intelligence (AI) Enhancement of Video or Augmented Reality (AR) Presentations of Live Events." It was filed on April 29, 2025, and published on October 29, 2026.

The wording matters. A published application is not an issued patent. One outlet says the patent office has approved it and that it will issue once Microsoft pays the final fee. Futurism reports the patent office has approved the application, and that once Microsoft pays the final fee it should issue as a granted patent. That claim is secondhand. The application record I reviewed shows a published application. So "filed" is the safe verb, and "patented" is premature.

Computerworld's brief credits Futurism, which credits the patent tracker Patentlyze. Computerworld also says it is unclear when, or whether, Microsoft will launch the technology. Futurism said Microsoft did not immediately respond to a request for comment.

How the system is described​

The filing covers two paths.

Post-production. If a script or transcript exists, a generative AI system edits it and swaps the matching words in the audio. It then generates transitions and adjusts the video to fit.

Near-real-time. An AI agent scans the live audio and picks out words to change. It uses a set of parameters and, in some embodiments, a viewer's profile. A generative AI system then replaces those words in both the audio and the video. It also edits the material just before and after each replacement so the join is smooth. The result is shown to the viewer in near-real-time.

The application defines near-real-time as within a few seconds, and gives 1–10 seconds as an example. It also ties that window to the delay already used in live broadcasts to let people filter profanity by hand. That is a described target. It is not a measured result.

The video edits could include changing the speaker's lips so they appear to say the replacement words. The filing says the processing could run on device CPUs, GPUs or NPUs. The listed devices include AR eyewear, smartphones, tablets and video-recording devices. Those are implementation options, not a product list.

What it could change in speech​

The filing lists these targets:

• Filler words such as "um," "uh," "like" and "you know."
• Personal words or phrases, including coined terms.
• Colloquialisms such as "y'all," "gonna" and "bite the bullet."
• Mumbled, mispronounced, misused or misspoken words.
• Offensive, insensitive or inappropriate words.
• Signs of a speech impediment, such as a stutter or lisp.

It also covers translation. Speech in a language outside the viewer's profile could be replaced with translated words, with the video changed to match. The scenarios in the claims include international meetings, podcasts, movies, opera and lectures.

The filing's own benefits section goes further than cleanup. It mentions "instant fact-checking" and live correction of inaccurate or misleading information. It also says replacement words could come from an internet or database search. Those are patent-described possibilities. Nothing shows they have been built or validated.

Viewer profiles​

The application's examples of profile data are:

• Age.
• Spoken languages.
• Region of birth or residence.
• Ethnicity.
• Cultural background.
• Religion.
• Education level.
• College major.
• Speech or video samples, with metadata.

This is a list of what a profile could contain. It does not show that any Microsoft product collects these attributes. Still, any system that edits speech per viewer would have to decide what it knows about that viewer.

Settings it could apply to​

The claims name a wide range of live content:

• Lectures and presentations.
• Social media videos and podcasts.
• Movies and television.
• Concerts, theater and opera.
• Awards shows, game shows and cooking shows.
• Video-game cut scenes and cloud-game streams.
• Sports.
• Training videos.

The "ground truth" provision​

The filing says the original live audio and video, as recorded, may be stored as "ground truth" for future reference. Futurism reads that as a nod to potential misuse. The language is optional. The application does not say viewers would be told a feed had been altered. It also does not require originals to be kept, or require any audit trail.

Analysis: translation is not substitution​

This is my own reading, not a claim from the filing. The proposal bundles several different things:

• Translation and removing "um" are accessibility and polish features. Many viewers would welcome them.
• Replacing colloquialisms erases how a person actually talks. Futurism's example is a Southerner's "y'all."
• Filtering offensive words is already done with broadcast delays. Here it becomes personal and visual, so the speaker's lips are altered too.
• Substituting or "fact-correcting" substantive statements changes meaning. That is a different kind of act from the other three.

The filing does not show how reliably such edits would preserve meaning. It does not show how users would choose settings, or who would decide what counts as "inappropriate" or "misleading." Those gaps are why the "ground truth" language matters.

Commentators have raised deepfake concerns. Mezha's coverage says the technology could enable real-time deepfakes. WION and Digit also flag misuse and authenticity worries. Those are concerns about what the design could enable. They are not documented incidents.

There is a counterargument. The patent frames the work as a fix for the "exact copy" model of recording. That model captures filler words and verbal tics, and removing them takes skilled technicians. Teams that make training videos or lectures could reasonably want cleanup tools. Seen that way, much of the filing is aimed at editing, not deception.

Related Microsoft work​

Tech Times, quoted by ISPR, notes that Microsoft published VASA-1 research in April 2024. That project showed AI-generated talking-face video from a still image and audio. The same coverage says it was separate research and does not show the live-broadcast system has been implemented.

What IT readers should take from this​

• Treat it as a signal, not a roadmap. Companies file many patents that never become products.
• Plan for provenance. If something like this ships in conferencing, streaming or training tools, admins will need policies on disclosure and on keeping unedited originals. That applies most to compliance recordings, meetings and legal or regulatory content.
• Watch the hardware angle. The NPU mention fits the broader push toward on-device AI in PCs. It is only a described option here.
• Ask about labeling. If a product ever appears, the first questions are whether altered feeds are labeled, whether the original is retained, and who controls the settings.

Until Microsoft says more, the filing shows what the company has considered, not what it will build.