Imagine watching a livestream and hearing a speaker say something they never actually said — not because the creator edited the replay, but because artificial intelligence quietly changed the words while the event was still happening. That is the unsettling possibility behind a newly published Microsoft patent application describing an AI system that can replace spoken words in live video and adjust the speaker’s mouth movements to match.

The proposal is not limited to removing profanity or cleaning up awkward pauses. The patent describes a system that could filter filler words, change slang and regional expressions, correct misspoken language, and even translate speech into another language while modifying the video so the speaker appears to say the replacement words. It could make live content easier to follow — but it also raises a serious question: if AI can rewrite a live broadcast, how will viewers know what was really said?

What Microsoft’s AI livestream patent actually describes

The application is titled Near Real-Time Artificial Intelligence (AI) Enhancement of Video or Augmented Reality (AR) Presentations of Live Events. It was filed by Microsoft Technology Licensing, LLC on April 29, 2025, and published on October 29, 2026, under U.S. publication number US 2026/0341309 A1 (application 19/193,093).

At its core, the system would analyze a live audio and video feed, identify words that meet selected rules or user-profile settings, and generate replacement audio and video. The modified version would then be delivered to the viewer in what Microsoft calls “near-real-time.” The patent defines that timing as within a few seconds — giving examples of roughly one to 10 seconds, or a delay comparable to a live broadcast buffer.

The patent discusses possible use across lectures, presentations, social-media videos, podcasts, concerts, television productions, award shows, cooking and game shows, and sporting events. It also describes devices such as smartphones, tablets, video cameras and augmented-reality eyewear as possible ways to run or access the system.

That distinction matters: this is a patent application describing a possible technology, not an announcement that Microsoft has launched a livestream-rewriting product or that platforms are already using it.

How the AI could change a speaker’s words — and their mouth movements

The process starts with the system scanning the speaker’s live audio and video. An AI agent identifies words or sounds that should be changed according to selected parameters and a viewer’s profile. A generative AI system then creates replacement words and edits the corresponding audio and video segments.

Here is the part that makes the proposal especially striking: the video would be modified to match the new audio. The patent describes generating replacement frames and, in some examples, adjusting the speaker’s lips so the person appears to be forming the substituted words. It also calls for smoothing the audio and video around the edited sections so the changes blend into the broadcast.

In other words, this is more than a conventional profanity delay or a muted word. The concept is to deliver a version of the event in which the sound and the visible speech have both been altered. If the system works as described, the viewer could see a speaker’s mouth move in sync with a replacement phrase rather than the words originally spoken.

Viewer profiles could decide which words get changed

The patent’s personalization element is one of its most consequential details. The system could use parameters and information in a user profile to decide which words to replace. The filing lists possible profile information such as age, spoken languages, region of birth or residence, cultural background, education, speech samples and video samples. Some of those examples include highly personal categories, although the filing does not establish that Microsoft is currently collecting or using such information for this purpose.

The words identified for possible removal or replacement go well beyond profanity. The patent gives examples including “um,” “uh,” “you know,” “basically,” and other filler words. It also mentions personal expressions, slang and colloquial phrases such as “gonna,” “gimme,” “y’all,” “ain’t,” and “bite the bullet.” Other listed categories include mumbled, mispronounced, misused or misspoken words, as well as offensive, insensitive or inappropriate language.

That could create a very different viewing experience from one person to another. One viewer might prefer to hear a speaker exactly as they sound. Another might want profanity filtered. A viewer who speaks a different language could receive translated speech. The patent even discusses using replacement words sourced through an internet search, opening the door to edits that do more than simply remove a sound.

Live translation is another major use Microsoft outlines

Not every proposed use is about censorship or polishing a speaker’s delivery. A separate claim in the application describes translating live speech into a language listed in the viewer’s profile. The system would identify words spoken in another language, replace them with translated words in the audio, and generate video frames that match the translated speech — including lip movements.

That could be useful for international meetings, classes, travel, livestreams, podcasts, films, television, concerts and other events where language differences keep people from following along. Instead of hearing a separate voiceover that does not match the person on screen, a viewer could potentially see the speaker’s mouth synchronized with the translated words.

The translation concept is distinct from the broader speech-filtering idea in the patent. Both involve near-real-time changes to audio and video, but the translation claim specifically focuses on replacing words from a language outside the viewer’s profile with translated words in a language the viewer understands.

The biggest concern: viewers may not know what was really said

The obvious benefit is convenience: cleaner presentations, fewer distracting filler words, language translation and potentially more accessible broadcasts. But the same technology could also change the meaning or character of what an audience sees and hears.

Removing a repeated “um” is not the same as replacing a speaker’s slang, changing a controversial phrase or substituting different words. Regional expressions such as “y’all” are part of how people communicate and where they come from. If a personalized filter treats those expressions as undesirable, it could make a speaker sound less like themselves — even when the original words were perfectly understandable.

There is also a question of accountability. A viewer could come away believing a person made a statement that was actually generated by an AI system. A replacement phrase could soften a criticism, remove a qualification, change a joke or alter the apparent meaning of a statement. If different viewers receive different versions of the same event, they may not even realize they are disagreeing about edited versions of the same broadcast.

The application includes a provision describing the original live audio and video being stored as “ground truth” for future reference. That is an important safeguard to note, but it should not be mistaken for a guarantee that every implementation would preserve, expose or clearly label the original recording. The patent describes possible system designs; it does not establish what a finished consumer product’s transparency rules would be.

Is Microsoft actually releasing this technology?

No release has been announced in the patent materials or the reporting reviewed for this article. The document is a published U.S. patent application, not proof that a finished feature is available, that Microsoft has committed to building it, or that a launch date exists. Patent applications describe inventions and seek legal protection for claimed ideas; many do not become commercial products in the form originally described.

The filing does, however, put a specific vision on paper: AI could become an intermediary between a speaker and the audience, changing not just the sound of a broadcast but also the video that viewers see. That is a much bigger step than captions, noise cancellation or a traditional content delay.

If a system like this ever reaches the public, the details will matter. Viewers should be able to tell when speech has been altered, access the unmodified version when appropriate, and understand why a word was replaced. Without those safeguards, a tool designed to make livestreams more polished or accessible could also make it harder to know what actually happened.

For now, Microsoft’s patent is a proposal — not a product. But it raises a question worth watching as generative AI moves deeper into live media: when a livestream can be rewritten before it reaches the audience, who gets to decide which version of reality people hear?

Sources and further reading

• U.S. Patent Application 2026/0341309 A1 — Near Real-Time AI Enhancement of Video or AR Presentations of Live Events
• Futurism: Microsoft Patents an AI That Can Literally Put Words in People’s Mouths in Real Time
• TechSpot: Microsoft patent filing describes AI that rewrites what people say in livestreams in real time
• Computerworld: Microsoft files patent for AI that can alter what people say during live broadcasts