A short demo shown ahead of Adobe’s MAX Sneaks revealed a tool built to shift the tone and delivery of a voice-over without a fresh recording. A plain, matter-of-fact narration played on a loop. After the presenter pulled up a transcript, highlighted a sentence, and picked an emotion from a list of presets, the same audio snapped into a new mood. What had sounded flat became confident, then dropped down to a near whisper, all in a few seconds.
Adobe calls the feature Corrective AI. It is one of several experiments scheduled for display at the MAX Sneaks portion of the MAX Creativity Conference in Los Angeles. Sneaks serves as a testing ground for prototypes and future features; past demonstrations have moved into Adobe’s mainstream apps within a short span. The company tends to show multiple workflows during the showcase that hint at how creative tools might feel in the near term.
Earlier at MAX, Adobe rolled out generative speech for Firefly, a system that offers a handful of preset voices plus emotion tags that change inflection. Corrective AI applies that speech-shaping approach to recorded material instead of replacing human performances. Editors can touch up an existing take by editing its transcript and choosing a different expressive setting, which keeps the underlying performance while altering delivery.
Lee Brimelow guided a different demo that split a single audio file into separate components. The prototype, labeled Project Clean Take, is capped at five discrete tracks: spoken parts, ambient background, isolated sound effects, and similar layers. The AI’s separation felt uncannily precise in several examples. In one clip a bell from a drawbridge nearly drowned out a host. After the model did its work the bell was gone from the vocal track. The presenter then dialed the bell back into the mix by raising the level on the isolated effect.
A separate scenario centered on creators who film in public where licensed music plays nearby. Unlicensed songs can trigger automated copyright flags on platforms like YouTube, which means background music can be a legal and logistical headache. In a tidy demo, the AI split the music from the rest of the audio, swapped it for a similar piece from Adobe Stock, and applied processing that recreated the reverb and atmosphere of the original clip. The change required only a few clicks, leaving the rest of the soundtrack intact.
A common thread through these prototypes is practical problem solving for editors and content makers. The tools aim to restore damaged audio, remove distractions, or prevent the need for a costly rerecording session when a performance is salvageable. The company presented newer generative audio features alongside corrective ones. For sound designers the claim is that a model can scan footage, identify moments that would benefit from effects, and populate those moments with AI-generated sounds that are cleared for commercial use.
Oriol Nieto put that claim to the test with a short scene reel that contained several vignettes and a single voice-over track, but no manual sound design. The model parsed the footage into scenes, attached emotional tags, and supplied brief descriptions of each moment. Sound effects then appeared where the model thought they belonged. It produced an alarm clock sting for one shot and, in a lighter example, added the thunk of a car door closing when the main character — an octopus, in this case — climbed into a vehicle.
The results were uneven. The alarm lacked realism, and in a beat where two characters embraced the model layered in a cloth-rustling effect that felt off. Rather than dig into manual edits, the team used a conversational interface modeled on ChatGPT to request fixes. In a clip where the car provided no ambient engine noise, the presenter typed a short directive asking for a car sound. The interface found the correct segment, generated a suitable effect, and placed it on the timeline with accurate timing.
None of the prototypes are shipping as finished features yet. Adobe has a track record of turning Sneaks concepts into product features; Harmonize is a case in point. Shown at a Sneaks event last year, Harmonize automatically positions assets so color and lighting look consistent inside a scene and it later arrived inside Photoshop. Representatives suggested similar items from this set of demos could arrive in Adobe’s apps sometime in 2026.
The timing of these demonstrations comes within months of a major labor moment in the wider entertainment world. Video game voice actors concluded a nearly year-long strike that secured contractual protections around AI use: companies must obtain consent and provide disclosure agreements when a developer wants to recreate a voice actor’s voice or likeness with AI. Performers have been preparing for AI’s impact for some time. Tools that finesse existing recordings without fabricating new voices still alter the dynamics of production work, and they add to the list of ways AI is reshaping creative practice.