AI Stem Splitter — Separate Any Song Into Vocals, Drums, Bass & Other
Drop in a song and an AI model splits it into four stems — vocals, drums, bass and everything else — right here in your browser. No upload, no account, and unlike the big AI audio services, downloading your stems is free. Make a karaoke track, lift an acapella, or study the bass line: your music never leaves your device.
How to split a song into stems
- Drop your song — MP3, WAV, M4A, OGG or FLAC, up to 12 minutes. It is decoded on your device; nothing is uploaded.
- First visit only: the browser downloads the AI model (about 172 MB) and caches it. Every later visit starts instantly.
- The AI separates the track into vocals, drums, bass and other. On a modern laptop this runs faster than the song plays; on CPU-only devices it can take one to three times the song length.
- Mix and download — play the stems together, mute, solo and ride each volume, then download any stem as WAV or MP3, or grab the instrumental in one click.
What each stem contains
| Stem | What lands in it | Typical use |
|---|---|---|
| Vocals | Lead and backing vocals, spoken parts | Acapellas, remixes, vocal practice |
| Drums | Kick, snare, hats, cymbals, percussion | Sampling, drum transcription, play-alongs |
| Bass | Bass guitar, synth bass, 808s | Bass practice, low-end study |
| Other | Guitars, keys, synths, pads, strings | Karaoke beds (with drums + bass), chord study |
How it works — a real AI model, on your device
The splitter uses HTDemucs, Meta AI’s hybrid-transformer separation model (MIT licensed), exported to ONNX and executed with ONNX Runtime Web — WebGPU where your hardware offers it, multi-threaded WebAssembly everywhere else. There is no server in the loop: the same class of model the subscription services run in their clouds runs here on your own machine, which is exactly why the downloads can be free and why your unreleased tracks stay private. Separation quality is within a whisper of the paid leaders — independent leaderboards put the best open models under 1 dB behind them.
Honest limits: the one-time model download is about 172 MB; phones and low-memory devices can run out of room on long songs, so desktop is recommended; and very dense, reverb-heavy mixes can leave a trace of one stem inside another — that is true of every stem splitter, including the paid ones.
Frequently asked questions
Is my song uploaded to a server?
No. The AI model runs entirely in your browser on your own device. The only download is the model itself, once — your audio never leaves your device.
Is the stem splitter really free?
Yes. Splitting and downloading are free, with no account and no watermark. Unlike most AI audio services, the free tier here lets you actually download your stems — as WAV or MP3.
What stems do I get?
Four stems: vocals, drums, bass and other (guitars, keys, synths and everything else). Play them in the built-in mixer with mute, solo and volume per stem, download each stem separately, or download the instrumental (drums + bass + other) in one file.
Why does the first split take a while?
On your first visit the browser downloads the AI model (about 172 MB) and caches it, so later visits start instantly. The separation itself runs on your device: with a modern laptop GPU (WebGPU) a song splits faster than it plays; on CPU it can take one to three times the song length.
Does it work on phones?
Short songs can work on recent phones, but the model plus a full song needs a lot of memory — a desktop or laptop browser is strongly recommended for reliable results. Tracks are limited to 12 minutes.