At a glance
- Cost
- 100% free, no account
- Separates into
- Vocals, drums, bass and other instruments
- Model
- Meta HTDemucs v4 neural network
- Privacy
- Runs locally; nothing is uploaded
- Download as
- WAV or MP3 per stem, or one ZIP
- Speed
- Roughly 1–3 minutes per song with a modern graphics card
Splitting a song into stems
- Open the track you want to take apart: MP3, WAV, FLAC, M4A or OGG.
- Keep the AI engine selected for the best separation, or pick Instant for a quick, rougher preview.
- Start the split. The first use downloads the AI model once (about 180 MB), then processing begins with a live time estimate.
- Grab your stems one by one, or all four together in a single ZIP file.
The four stems explained
| Stem | You'll hear | Great for |
|---|---|---|
| Vocals | Lead voice, harmonies and their effects | Acapellas, remixes, lyric transcription |
| Drums | Kit and percussion | Drum practice, breakbeats, groove study |
| Bass | Bass guitar, sub-bass, 808s | Learning bass lines, re-amping |
| Other | Guitars, keys, synths, strings, pads | Chord study, backing tracks |
Because the stems are complementary, lining them up in any DAW at unity gain reproduces the original mix. That makes it easy to rebalance a song, for example turning the vocals down 6 dB for a sing-along version.
How the AI tells instruments apart
HTDemucs v4 (Hybrid Transformer Demucs) from Meta AI Research was trained on multitrack recordings where every instrument was available separately. It learned the fingerprints of each source: the sharp, noisy onset of a snare, the steady low fundamental of a bass, the formants of a human voice. When you give it a mix, it predicts each of those parts at once.
Your song is analysed in overlapping chunks of just under eight seconds, which are blended back together so the stems play continuously. The work happens on your graphics card via WebGPU when available, or on the CPU otherwise. Nothing leaves your computer.
Ways to use your stems
- Play-along tracks: drop the drum or bass stem and take over that part yourself.
- Transcription: solo the bass or "other" stem to hear notes hidden in a busy mix.
- Remixes and edits: build new arrangements in GarageBand, Ableton Live, FL Studio, Logic or Reaper.
- DJ tools: acapella intros and drum-only transitions.
- Change the key: re-mix the non-vocal stems and transpose them with the Key Changer.
Stems remain the property of the original rights holders. Please get permission before sharing or releasing anything built from someone else's recording.
Common questions
Is the stem splitter really free?
Yes. There are no credits, subscriptions or length limits. Split as many songs as you like.
Are my songs uploaded for processing?
No. The AI model is downloaded into your browser and runs locally. Your audio never leaves your device.
How long does a split take?
On a recent graphics card, a typical 4-minute song takes about 1–3 minutes. Older integrated graphics and CPU-only computers can take considerably longer. A countdown shows the estimated time left.
Can I get guitar and piano as separate stems?
Not yet. Guitar and keyboard parts are delivered together in the "other" stem. We'll add dedicated stems when a browser-compatible model is available.
WAV or MP3: which should I choose?
WAV is lossless and best for further editing. MP3 is far smaller and fine for listening or practice.