How the stem splitter works
A stem splitter takes a finished stereo song and estimates what each instrument group sounded like on its own. AudioDrop does this in two passes so you never stare at a progress bar.
First, a fast spectral separation gives you a rough preview split in about a second. You can press play and start exploring right away. Then the AI model, HTDemucs (Hybrid Transformer Demucs, from Meta AI research), runs in the background and replaces the preview with a much cleaner result.
The model runs with WebAssembly across several CPU cores in parallel. On a modern laptop the AI pass runs faster than real time, roughly a 3-minute song in under a minute. Older machines and phones take longer.
What is in each stem
You get four stems. Here is what usually lands in each one.
| Stem | What it contains | Typical use |
|---|---|---|
| Vocals | Lead and backing vocals, spoken parts | Acapellas, remixes, vocal study |
| Drums | Kick, snare, hats, cymbals, percussion | Sampling, drum transcription |
| Bass | Bass guitar, synth bass, 808s | Bass practice, low-end study |
| Other | Guitars, keys, synths, pads, strings | Instrumental beds, chord study |
What you can do with split stems
Once a song is split you can recombine the stems any way you like. A few common jobs:
- Make an instrumental by muting vocals.
- Pull an acapella for a remix or mashup by soloing vocals.
- Practise drums, bass or guitar by muting your own part and playing along.
- Transcribe a busy part by soloing it so you can hear every note.
- Study a mix by listening to how each element sits on its own.
- Build a DJ edit by dropping the drums out for a breakdown.
Tips for the cleanest stem separation
Start from the best source you have. A WAV or FLAC file, or a high bitrate MP3, gives the AI model more detail to work with than a low bitrate stream rip.
Wait for the AI pass before you download. The preview is for listening, the refined split is what you want to keep. Modern studio productions with clear, separate parts split best. Dense live recordings are harder.
If a stem sounds thin on its own, check it in context. Small artefacts often disappear once the other stems are back in the mix at a lower level.
Honest limitations
No stem splitter is perfect, and this one is no exception. You will hear some bleed between stems, for example a hint of hi-hat in the vocal stem. Reverb tails and backing vocals can land in odd places. Heavily distorted guitars, lo-fi recordings and old mono tapes separate less cleanly than modern productions.
The first split also needs the AI model, which is about 166 MB. It downloads once and is cached by your browser after that.
Preview split vs AI split
The two passes do different jobs. The preview gets you listening immediately. The AI split is the one you keep. You never have to choose, since the preview is replaced automatically as soon as the AI pass is done.
The first time you use the stem splitter, the browser also fetches the AI model, about 166 MB. That is a one-off. After it is cached, new songs go straight to separation, even after you close and reopen the page.
| Preview split | AI split (HTDemucs) | |
|---|---|---|
| Ready in | About a second | Faster than real time on a modern laptop |
| Method | Fast spectral separation | Hybrid Transformer Demucs model |
| Quality | Rough, more bleed | Much cleaner stems |
| Best for | Quick listening and checking | Downloading and real work |
| Needs model download | No | Once, about 166 MB, then cached |
A stem splitter with no upload
Most online stem splitters send your song to a server. AudioDrop runs the whole AI stem separation in your browser, so your file never leaves your device. There is no account, no queue, no watermark and no charge per song.
It works in desktop Chrome, Edge, Firefox and Safari. It also runs in mobile browsers on iPhone and Android, but the AI model is heavy on a phone, so a laptop is the better choice for long songs. If you publish anything made from someone else's music, make sure you have the rights.