How to get stems from an AI-generated track with AI stem separation

Suno and Udio hand you a finished master — vocals, drums, bass, and everything else baked into one stereo file. That is fine if you only want to listen. It is a problem the moment you want to remix the track, build a karaoke version for a cover, lift the vocal into a different beat, or pull just the spoken portion of a track for transcription. The fix is AI stem separation: a source-separation model that takes the mixed file and estimates each part of the mix on its own track.

What stem separation actually is

Stem separation is not unmixing. It is not reaching back into the session and pulling out the original recordings, because for an AI-generated track those recordings never existed. The model is making an educated guess. It listens to the mixed audio, recognises what a vocal sounds like, what a kick drum sounds like, what a bass line sounds like, and reconstructs an estimate of each as a separate stem.

That distinction matters because it sets expectations. A separated vocal will have a small amount of room reverb, maybe a hint of snare on loud hits, and a bit of bass rumble underneath. A separated drum stem might have a ghost of the vocal on long sustained notes. This is bleed. It is not a bug — it is what every source-separation model produces. The cleaner the source recording, the less bleed you get.

Why AI-generated mixes are a particular case

Generative audio is an interesting input for a separator. On one hand, the mixes are usually well-balanced — vocals sit forward, drums sit back, the bass holds its register — which gives the model strong cues to work with. On the other hand, the vocoder-style artifacts that show up in Suno and Udio output can blur transient edges and smear high-frequency detail. That makes the boundaries between sources slightly fuzzier than they would be on a studio recording, which can cause a touch more bleed.

The practical result: AI stems are typically workable, sometimes excellent, occasionally rough — depending on the track. Acoustic-leaning material with clear instrumentation tends to split cleanly. Dense electronic mixes with a lot of layered synths in the same register are the hardest, because the model has fewer cues to distinguish parts.

The three engines and when to use each

The Refiner exposes three separation engines. Each one costs 3 credits per split regardless of which you pick, so the choice is purely about which output suits the job.

Demucs is a classic source-separation model. It outputs four stems — vocals, drums, bass, and other (everything that was not a vocal, drum, or bass). It is fast and produces solid results on most material, which makes it the right default. If you are sketching out a remix idea, building a quick karaoke version, or deciding whether stems on this particular track are usable at all, start here.

Mel-Band (Mel-Band RoFormer) is a newer transformer architecture that produces the same four-stem output as Demucs with less bleed between stems. It takes longer to run. Use it when stem quality matters more than speed — a final remix you are mixing into a DAW, a vocal you want to sample prominently, or anything that will be scrutinised on headphones.

Voice Cleanup is a two-stem split designed for speech rather than music. It separates the speech track from everything else — music, room tone, applause, traffic. That makes it the right tool for podcasts, interviews, voice memos, and prepping audio for transcription. It is not the right tool for separating a sung vocal from a band; reach for Demucs or Mel-Band there.

The workflow, including the optional refine step

Drop your track into the stems flow, pick an engine, and run the split. Tracks up to 8 minutes are supported. A typical 3-minute song finishes in roughly 1–2 minutes depending on the engine. When the run completes, the in-app multitrack player lets you preview the stems side-by-side before downloading. The output is one WAV per stem plus a ZIP with all of them, so you can grab everything in a single click.

If your source has strong vocoder artifacts and the first split sounds rougher than you want, the legitimate next move is to refine the source first and then split that. Refinement is a separate operation — there is no combined refine-and-split button. You run Enhanced or Flow Matching first, download the refined WAV, and upload that to the stems flow. Our guide on Enhanced vs Flow Matching covers which refinement preset to pick.

What to actually do with the stems

A few concrete workflows people run on AI stems:

DAW remix. Drop the four stems into your DAW as separate tracks. Now you can swap the drums for live samples, retune the bass, sidechain things to the kick, or rebuild the arrangement. The "other" stem will carry pads and synths from the original mix — sometimes you keep it, sometimes you replace it.

Karaoke / instrumental. Take the bundle, mute the vocal stem, and bounce the rest down to a single instrumental track. Useful for covers, lyric videos, or singing over your own AI-generated backing.

Isolated vocal. The inverse — keep only the vocal stem for use over a different beat, or as a sample. Mel-Band tends to be the better engine here because the vocal is what you are going to scrutinise.

Speech transcription. Voice Cleanup on a podcast or interview recording strips background music and room noise, which meaningfully improves transcription accuracy on Whisper or similar models. It is also handy for cleaning up voice memos before you send them.

Honest limits

Stems will always have some bleed. The cleaner the input, the cleaner the split — which is why the optional refine-first step exists. We ship four-stem separation (or two for Voice Cleanup); we do not currently split a drum stem further into kick, snare, and hi-hat, and we do not promise specific accuracy numbers because they vary too much track to track. Tracks must be under 8 minutes, same as refinement.

Stem separation is a Pro feature on The Refiner. Free accounts get one lifetime trial split so you can run the engines on your own audio before deciding. After that, splits cost 3 credits each and are included in the monthly allowance on Pro and Studio plans. The full breakdown is on the pricing page, and the engine reference lives in /help#stems. Start free, run the trial split on a Suno or Udio track you already have, and you will know within a minute whether the stems are clean enough for what you want to do with them. The fastest sanity check on whether refinement helps is to split once before and once after — the difference is usually audible on the vocal. Demos comparing the engines live on the homepage.