To split a song into stems, run the cleanest file you have through an AI separation model. Demucs-based tools give the best offline quality, and Moises, lalal.ai, RipX, and Veena all wrap that kind of model in something you can actually use. Expect two, four, or six stems, expect artifacts on dense modern masters, and plan to repair the result rather than expecting a perfect extraction.
- The cleanest source file you can get — WAV or FLAC beats a streaming rip
- A separation tool (browser, desktop app, or command line)
- A DAW or editor to audition and repair the stems
- Around five to fifteen minutes for a typical song
Separation quality is capped by source quality. Lossy encoding throws away information the model needs to tell one instrument from another, and it cannot be recovered.
Use the purchased download or the original master if you have it. If your only copy is a 128kbps MP3, separation will still work, but the high end will sound papery. Never re-encode before separating.
Most tools offer 2-stem (vocal and instrumental), 4-stem (vocals, drums, bass, other), or 6-stem (adding guitar and piano). Every extra stem is another decision the model has to guess at, so ask for the fewest stems your project actually needs.
Making an instrumental? Ask for 2. Remixing? Ask for 4.
| Tool | Where it runs | Best for |
|---|
| Demucs | Command line, open source, free | Maximum control and no cost, if you are comfortable with a terminal |
| Moises | Browser and mobile | Quick 4-stem splits, practice, transcription |
| lalal.ai | Browser | Occasional one-off jobs |
| RipX | Desktop | Note-level editing after separation |
| Veena | Desktop browser | Separating and then immediately rebuilding in the same project |
Veena runs separation in the browser and drops the stems straight onto a multitrack timeline, which matters because the next step after separating is almost always editing. Its real limit is that it needs a desktop browser and a connection — it is not a phone tool.
Listen to each stem on its own, at the loudest chorus and at the quietest verse. Separation fails differently in each. What you are listening for:
- Bleed — bass energy sitting in the drum stem, or a ghost vocal in the other stem
- Watery or phasey vocals — the classic sign of masking errors
- Missing cymbals and sibilance — high-frequency content that overlapped and got split badly
- Smeared reverb tails — reverb that belonged to one source spread across several
Fix in this order, gently:
- High-pass the drum stem around 60-80Hz if bass has leaked into it
- Use a de-esser rather than a wide EQ cut for harsh separated vocals
- Add a small amount of new reverb to a stem that sounds dry and stripped — it masks smearing
- Where the drum stem is unusable, program a replacement kit instead of fighting it. Replacing usually beats repairing.
Do not reach for aggressive noise reduction. It trades one artifact for another.
Import all stems, set them to unity gain, and compare the sum against the original. It should be close in level and tone. If it is dramatically different, the export gain staging is wrong somewhere.
Sparse, dynamic, wide mixes separate well. Dense, brickwalled, mono-heavy masters separate worst, because loudness limiting compresses away the differences the model uses to tell sources apart. A 1970s soul record usually separates better than a modern loudness-war master.
Separated stems are a reconstruction, not a recovery. Treat them as raw material.
Related reading: how stem separation works, best stem separation tools, and AI music and sample clearance.
Start making music in Veena
Free, browser-based, no downloads required.
Try Veena Free