A free AI stem splitter that runs on your own computer
A stem splitter takes a finished, mixed song and pulls it apart into the parts it was built from. Producers call those parts stems. Most online stem splitters do this on a server: you upload the track, wait in a queue, and then pay or watch a credit counter once the free minutes run out. This stem splitter works differently. The AI model is downloaded into your browser once and does all the work on your own graphics card, so your music stays on your machine and you can split as many songs as you want.
Under the hood it is the same kind of AI stem separator that powers paid apps: HT-Demucs v4, the hybrid transformer model released by Meta AI Research under the MIT license. We run it with ONNX Runtime and WebGPU, which lets a web page use the GPU almost as directly as a desktop program.
- Vocals
- Drums
- Bass
- Other
How to use the stem splitter
- Drop in a song. Drag an MP3, WAV, FLAC, M4A or OGG file onto the box above, or click to choose one. Video files such as MP4 work too; the stem splitter only reads their audio track.
- Wait for the model on the first run. The AI model is about 170 MB. It downloads once and is stored in your browser cache, so the next song starts in about a second.
- Watch the progress bar. The song is processed in overlapping 8-second segments. The bar shows how much is done and roughly how long is left.
- Listen and mix. When it finishes, press play. Every stem has mute (M), solo (S) and a volume slider, so you can hear the drums alone, drop the vocals, or turn the bass up.
- Download. Save each stem as a WAV file, grab all four stems in one ZIP, or download the mix exactly as you set it.
What each stem contains
| Stem | What ends up in it | Typical uses |
|---|---|---|
| Vocals | Lead vocal, backing vocals, harmonies, ad-libs, rap, spoken word | Acapellas for remixes, vocal practice, transcribing lyrics |
| Drums | Kick, snare, hi-hats, cymbals, toms and most percussion | Studying grooves, sampling breaks, drum practice |
| Bass | Bass guitar, synth bass, sub bass, 808s | Learning bass lines, play-along tracks |
| Other | Guitars, piano, keys, synths, strings, brass, pads, effects | Chord study, instrumental beds, sampling |
Add the four stems back together and you get the original song again, almost exactly. That is a useful check: if a sound is missing from one stem, it is sitting in another one. The stem splitter never throws audio away; it only decides which stem each sound belongs to.
How good is the separation?
Stem separation quality is usually measured in SDR (signal-to-distortion ratio) on the MUSDB18 test set. HT-Demucs reaches about 9 dB averaged across the four stems in the paper by Rouard, Massa and Défossez (ICASSP 2023), one of the best published results for an open model. In plain words, vocals and drums come out clean enough for remixing on most pop, rock, hip-hop and electronic songs. You will still hear some bleed in hard cases:
- Heavy reverb and delay on the vocal often stays partly in the other stem.
- Distorted guitars and bass that share the same frequencies can swap small pieces.
- Live and old mono recordings separate less cleanly than modern studio mixes.
- Instruments the model was not trained on, such as accordion or sitar, usually land in other.
We checked the browser version against the original PyTorch model on the same audio. The stems match to within a fraction of a decibel, so you get the full model, not a cut-down version.
Why use a stem splitter in the browser?
Privacy
Your song is never uploaded. That matters for unreleased demos, client mixes, rehearsal recordings and anything under NDA. You can even disconnect from the internet after the model has loaded and the stem splitter keeps working.
No limits and no sign-up
There is no account, no credit system, no file count and no watermark. Long DJ edits and full albums are limited only by your computer's memory; the tool accepts tracks up to 10 minutes each.
Speed
There is no upload and no queue. On a recent graphics card the stem splitter processes a song several times faster than real time. In our tests an RTX 4060 Ti split a four-minute song in about 22 seconds.
Who uses a stem splitter
- Producers and DJs pull acapellas and instrumentals for remixes, bootlegs, mashups and live edits.
- Musicians mute their own instrument and play along, or solo it to learn a part by ear.
- Singers and karaoke fans remove the vocals to sing over the backing track.
- Teachers and students isolate a bass line or a drum pattern to study arrangement and groove.
- Video editors and podcasters lower the music under dialogue, or keep only the beat of a song.
Stem splitter vs vocal remover
A vocal remover gives you two tracks: the voice and everything else. A stem splitter goes further and also separates drums, bass and the remaining instruments. Because this tool always produces all four stems, it covers both jobs. Mute the vocals and you have an instrumental; solo them and you have an acapella. For one-click versions of those jobs, see the vocal remover, the acapella extractor and the drum remover.
Tips for the cleanest stems
- Use the best source you have. A lossless WAV or FLAC, or a 320 kbps MP3, separates more cleanly than a low-bitrate stream rip.
- Split the full song, not a short loop. The model hears context, and very short clips can sound thinner.
- Keep the browser tab in front while it runs. Background tabs can be slowed down by the operating system.
- If you only need two tracks, still split into four stems. You can mix any combination afterwards without running it again.
Browser and device requirements
The fast path needs WebGPU, which current Chrome and Edge support on Windows, macOS and ChromeOS. Recent Safari and Firefox versions are adding it too. A graphics card with 4 GB of memory or more is comfortable. Without WebGPU the stem splitter still works on the CPU through WebAssembly, only much more slowly, and the page tells you when that happens. Expect the tab to use one to two gigabytes of memory for a typical song.