Takeaway: On October 29, 2026, the U.S. Patent and Trademark Office published Microsoft patent application US 2026/0341309 A1 (application number 20260341309), titled “Near Real-Time Artificial Intelligence (AI) Enhancement of Video or Augmented Reality (AR) Presentations of Live Events.” The filing, submitted April 29, 2025, describes software that scans live video and audio of people speaking, picks words to change using rules and a viewer’s profile, replaces those words in both the soundtrack and the picture, redraws surrounding frames — including the speaker’s lips — and presents the result within a few seconds. The application defines “near-real-time” as roughly 1 to 10 seconds, or a delay similar to the broadcast buffer already used to bleep profanity by hand. Suggested uses in the specification include cleaning filler words, translating speech with matching lip motion, and what the text calls “instant fact-checking,” with replacement words that can come from an internet or database search. This is a published application, not a granted patent and not a shipping Microsoft product. Patent trackers report that an examiner rejected the claims as filed on July 29, 2026, a routine office action that Microsoft can answer with amendments.
This is a trust-and-interfaces story about what it would mean if a livestream no longer guaranteed that the mouth you see matches the words that were actually said. It is not a product launch, not a Teams feature announcement, and not evidence that Microsoft is about to rewrite live news or video calls. Microsoft’s existing Interpreter capability in Teams, as secondary coverage has noted, translates audio and does not alter a speaker’s face. The filing is also distinct from debates about how models are trained on public web text, including the shared robots.txt AI-training preference token and practical crawler allow/block choices. Those fights are about what goes into a model. This application is about what a model could do to a live person on the way out to a viewer.
What the application actually describes
The published text covers two related systems. One is post-production: take an existing video and audio file, optionally compare it with a script or transcript, replace designated words, generate smooth audio transitions, and modify the matching video frames. The other — and the part that has drawn attention since patent trackers and trade sites surfaced the publication in early November — is near-real-time enhancement of a live event or an augmented-reality presentation of one.
Independent claim 1, as published on Justia’s copy of the application, recites a system that: scans a live video recording and corresponding live audio of people speaking, using an AI agent; identifies spoken words to modify “according to a first set of parameters and based on a profile of a user”; has a generative AI system replace those words in both the video and the audio, in near-real-time, based on that profile; generates “video and audio smooth transitions” before and after each replacement; and presents the modified portions to the user in near-real-time. A dependent path in claim 2 makes the video step more explicit: generate replacement video frames “to match the one or more replacement words in the live audio recording and the audio smooth transitions, using a video rendering system.” Elsewhere the specification says those replacement frames can modify the lips of the speaker so they appear to be saying the new words.
Claim 1 does not limit why a word gets changed. The specification fills that gap with a long list of “designated words”: filler words such as “um,” “uh,” “like,” and “you know”; personal turns of phrase; colloquialisms the text itself examples with “y’all,” “gonna,” and “bite the bullet”; mumbled, mispronounced, misused, or misspoken words; and “offensive,” “insensitive,” or “inappropriate” words. Non-verbal clean-up is also in scope — verbal pauses over a threshold, background noise, audio feedback. Replacement text, the claims say, can be retrieved from “an Internet search or a database search of general information regarding a topic of discussion.” That last clause is what lets secondary coverage describe the system as capable of more than tidying a presentation.
How “near-real-time” is defined — and what a profile can hold
The specification is unusually concrete about delay. Near-real-time “refers to within a few seconds (e.g., 1-10 seconds) or a delay that is used in modern live broadcasts (also referred to as a ‘live feed buffer’) to allow manual adjustments, e.g., filtering out profanities.” That is an example window, not a measured prototype result. It is also an admission that the edited stream is not simultaneous with the event. Someone in the room and someone watching through the system would not hear the same thing at the same moment, and two remote viewers need not hear the same thing as each other.
The viewer profile is the other load-bearing idea. Claim 5 says a profile may include age, spoken languages, region of birth, region of residence, ethnicity, cultural background, religion, education level, college major, and speech or video samples of the user (or, in the broader specification, of a speaker or intended audience member). Those fields describe what a profile could contain in the invention as claimed. They are not a list of data any current Microsoft product collects for this purpose, and the application does not say a commercial service would require all of them. The design implication is still sharp: the same live event could be rewritten differently for a child, a speaker of another language, or a viewer in another region, and claim 1 does not require the system to disclose that an edit occurred.
The devices listed are personal, not only studio hardware. Claim 3 says the system may be “a set of augmented reality (‘AR’) eyewear with at least one connected audio output device, a smart phone, a tablet computer, or a video recording device,” running the AI agent on NPUs, GPUs, and/or CPUs on the device. If that architecture were built, the rewrite would not have to happen in a broadcaster’s control room. It could happen on the viewer’s glasses.
Translation, “fact-checking,” and the ground-truth claim
Live translation is the use case that is easiest to defend on its face. Claim 7 (depending from the profile language in claims 5 and 6) and a separate independent system claim later in the set describe detecting words in a language that is not among the languages stored in the viewer’s profile, replacing them with translated words, generating smooth audio transitions, and producing replacement video frames “including rendering of lips of the one or more individuals to match the one or more translated words.” Example settings in the text range from international meetings and tourist visits to lectures, concerts, awards shows, game shows, and live sports. YouTube and other platforms already offer AI audio translation; the distinctive claim here is matching lip motion so the translated soundtrack does not fight the picture.
The more contested passage sits in the detailed description of the live path. The application says “near-instantaneous or live correction of inaccurate or misleading information by a speaker (e.g., instant fact-checking) may be achieved by using AI systems to search (on the Internet or other data sources) and verify, in near-real-time, information being spoken by the speaker(s).” Combined with claim 11’s permission to pull replacement words from an internet or database search, that is a design in which a model could overwrite a live sentence with text it retrieved moments earlier, then paint the speaker’s mouth to match. The filing does not define an authority for what counts as a “fact,” a user-visible diff, or an appeal path if the retrieval is wrong.
One dependent claim does preserve an original. Claim 14, depending from claim 1, states that “the live video recording and the corresponding live audio recording, as originally recorded, are stored as ground truth for future reference.” That is useful archival language. It is not the same as showing the viewer a label, a highlight, or a toggle that reveals what changed. Ground truth stored somewhere is not ground truth presented to the person who is being persuaded by the edited face. Secondary write-ups from Patentlyze, TechSpot, and Tbreak all note the same gap: the independent live-rewrite claim is silent on notifying the audience.
The live video recording and the corresponding live audio recording, as originally recorded, are stored as ground truth for future reference.
US 2026/0341309 A1, claim 14 (published October 29, 2026)
What the legal status actually is
A published application is a public disclosure of what the applicant hopes to own, not a government endorsement that the idea works or that the claims are patentable as written. Microsoft Technology Licensing is listed as the applicant on the Justia record. Coverage that surfaced the filing in early November, including Patentlyze, TechSpot, Tbreak, and TechPowerUp, has been careful to call it a filing rather than a product.
Patentlyze recorded that an examiner raised objections on July 29, 2026, and that the company can respond. Tbreak likewise reported that the examiner rejected the claims as filed on that date and described the step as routine, usually followed by amendments. First office actions that reject claims as filed are common, especially when independent claim 1 is written this broadly. A claim that wide invites narrowing. Until Microsoft responds and the examiner allows something — or the application goes abandoned — there is no issued U.S. patent on these claims.
Microsoft, in the secondary accounts we reviewed, has not confirmed plans to ship the system. Tbreak reported that the company did not respond to a request for comment from Futurism. Absence of comment is not confirmation and not a denial. It is simply where the public record stands as of this writing.
Why the design worries people who work in live media
The practical uses are real. Cleaning “ums” out of a lecture, translating a keynote so lips match, or giving a presenter with a stutter the option to smooth a recording are accessibility and production problems that editors already solve offline with expensive tools. The application’s own background section says current capture follows an “exact copy” model and that removing filler words later requires skilled technicians. Moving that cleanup into a few seconds of buffer is, in isolation, a productivity claim.
The trust problem is that the same pipeline does not have to stop at filler words. A system that can replace “offensive” or “inappropriate” language according to a viewer profile, pull substitute wording from a search index, and redraw a face has the affordances of a personalized deepfake applied to someone who is still speaking. Two viewers of the “same” livestream could receive different sentences. A speaker could be made to appear to correct a statement they did not correct, or to retract a claim they were in the middle of defending. Because claim 1 does not require disclosure, the edited version can be the only version a viewer ever sees.
That is a different failure mode from ordinary broadcast delay. A seven-second profanity dump removes a word; it does not invent a better one and paint it onto a mouth. It is also different from burned-in captions or a dubbed soundtrack that the audience can mentally discount. Lip-synced generative replacement is designed to defeat the visual check people still use when they suspect a voice-over. TechSpot’s November 7 note put the point bluntly: real-time face swapping is already popular with scammers, and altering what is being said in a livestream would raise the ceiling on both fraud and well-intentioned “correction.” Those are risks about what the design could enable. There are no documented incidents involving this application, because it is not a released product.
Microsoft is not alone in patenting live AI edits. Secondary coverage has pointed to Sony filings that describe pausing, blurring, muting, or replacing sensitive moments in games under parent-set rules, and to other Microsoft applications around generative alteration of interactive content. A cluster of filings is evidence of research interest, not of a ship date. For operators who already worry about synthetic media, the useful question is narrower: if tools like this land in consumer devices, how will platforms label an altered live feed, and will the original “ground truth” in claim 14 be available to the viewer, to the speaker, to a court, or only to the company that ran the model?
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| US 2026/0341309 A1 was published on October 29, 2026 | Confirmed | USPTO publication of application 20260341309; filed April 29, 2025 |
| Title covers near-real-time AI enhancement of live video or AR presentations | Confirmed | Also describes a post-production path using scripts or transcripts |
| The system rewrites selected spoken words and can redraw lips to match | Described in the specification and claims | Claim 1 covers audio and video replacement; claim 2 and the spec mention replacement frames / lips |
| Near-real-time means about 1 to 10 seconds | Example in the specification | Tied to existing live-broadcast delay for manual profanity filtering; not a benchmark result |
| Edits can be driven by a viewer profile (age, language, region, and other fields) | Described | A list of what a profile could include, not data any current product is shown to collect |
| Replacement words may come from an internet or database search; “instant fact-checking” is mentioned | Described | Claim 11 and the detailed description; no standard of truth is defined |
| The original recording is kept as “ground truth” | Dependent claim 14 only | Does not require notifying the viewer that an edit occurred |
| This is a granted Microsoft patent or a shipping feature | No | Published application only |
| An examiner rejected the claims as filed on July 29, 2026 | Reported by patent trackers | Described as a routine first office action; Microsoft can amend |
| Microsoft has announced plans to ship this | No | No product confirmation in the coverage we reviewed |
FAQ
Is US 2026/0341309 A1 a granted patent?
No. It is a published U.S. patent application. The USPTO made the filing public on October 29, 2026. Publication tells competitors what Microsoft has asked to protect. It does not mean the claims have been allowed or that a product exists.
When was the application filed, and what is its status?
It was filed on April 29, 2025, and published on October 29, 2026. Patent-tracking coverage reports that an examiner rejected the claims as filed on July 29, 2026. That kind of first rejection is common; the applicant can respond with amendments. There is no issued patent on these claims yet.
Would viewers be told that their livestream had been rewritten?
Independent claim 1 does not require any notice. Dependent claim 14 says the original video and audio “as originally recorded” may be stored as ground truth for later reference. Storage is not the same as a label, a highlight, or a side-by-side view for the audience.
Could two people watching the same event hear different words?
Yes, under the design as claimed. Words are selected using a set of parameters and a user profile that can include age, languages, region, and other attributes. Two profiles could trigger two different replacement sets, each with matching lip motion.
Does this mean Microsoft will rewrite live news or Teams meetings?
No public announcement says so. The application lists lectures, sports, awards shows, and similar live settings as examples, and it names phones, tablets, cameras, and AR glasses as possible devices. Microsoft’s shipping Interpreter feature in Teams, per secondary coverage, translates audio without changing the speaker’s face. A patent application is not a roadmap.
What is the “instant fact-checking” piece?
The detailed description says AI systems could search the internet or other data sources and verify information a speaker is saying in near-real-time, then correct “inaccurate or misleading” statements. Claim 11 allows replacement words to come from an internet or database search. The filing does not say who decides what is accurate, or whether the viewer would see the original sentence.
Why keep the original recording at all?
Claim 14’s “ground truth” language would matter for disputes, journalism, and any later audit of what a speaker actually said. It is one of the few built-in checks in the claim set. Whether that original would be available to the speaker, the viewer, a platform, or only the operator of the system is not specified in the published claims.
Sources: Justia — U.S. Patent Application 20260341309 (US 2026/0341309 A1, published October 29, 2026; filed April 29, 2025); Patentlyze — Microsoft Patent: Live AI Video That Rewrites Speech (updated October 30, 2026); TechSpot — Microsoft patent filing describes AI that rewrites what people say in livestreams (Daniel Sims, November 7, 2026); Tbreak — Microsoft patent describes AI that rewrites live speech and redraws lips (November 8, 2026). Quotations and claim numbers above are taken from the published application text. We have not treated secondary summaries as a substitute for the claims themselves.