Adobe uses AI to unmix messy audio into clean tracks

Adobe just revealed an experimental AI tool called Project Sound Lift, aiming to fix messy real-world recordings by tearing mixed audio tracks apart layer by layer.

Anyone who has recorded a podcast or shot video in the field knows the pain of an otherwise perfect take ruined by typing, sirens, or a burst of laughter. Adobe’s new prototype wants to turn that nightmare into a one-click fix.

Unmixing the unmixable

In a demo shared with The Verge, the workflow looks remarkably simple. You import a messy audio recording into the app and choose exactly what sounds you want to strip away.

Instead of just giving you a generic slider, the tool lets you target specific categories like applause, laughter, sirens, speech, crowd noise, typing, and traffic. Once processed, Project Sound Lift splits the audio into separate tracks, isolating the main voice or instrument from the background noise.

Editors can then bring those independent stems directly into Adobe Premiere Pro. That means you do not have to nuke the background entirely; you can simply lower the volume of the passing car while dialing in the clarity of the speaker.

Similar isolation tools exist today, such as Supertone Clear and Goyo voice separator, but they typically divide sound into blunt buckets like ambient noise, voice, and reverb. Adobe is trying to target specific sonic elements directly. It is the same flavor of machine learning technology that Peter Jackson and his team used to pull John Lennon’s vocals off a muddy cassette tape for the recent Beatles track, “Now and Then.”

Still, you cannot download it just yet. Adobe showed this off as one of its “Sneak” previews, meaning it is an early technology demonstration that may take a while to reach a public build, assuming Adobe ships it at all.

Why it matters

For indie filmmakers, podcasters, and solo YouTubers, audio is usually the hardest thing to control and the easiest thing to ruin a project. If Project Sound Lift works as advertised, it turns unusable field recordings into usable B-roll or interview tracks without requiring manual spectral editing in specialized software.

We think Adobe’s smartest decision here is splitting audio into stems rather than applying destructive noise suppression. Pure noise cancellation often strips out the life of a room, leaving voices sounding hollow and robotic. By letting editors retain the background noise on a separate track and tweak its volume, Adobe preserves natural room tone.

However, the catch with early audio demos is that they usually cherry-pick clean separations. In the real world, aggressive frequency isolation frequently leaves behind watery phase artifacts. Plus, because this is still labeled a “Sneak,” users should treat it as a peek at where Adobe wants Premiere Pro to go rather than a feature they can rely on this month.