How to split a song into stems in 4 steps
- 01
Load your track
Use Open file, drag a track into the drop zone, or select Try a demo to test the tool on a built-in loop first. Anything your browser can decode works, including MP3, WAV, FLAC, and M4A.
- 02
Choose quick preview or a full split
Turn on the 30-second quick preview to check separation quality fast, or run a full split of up to 10 minutes. The first split downloads the AI model once; it is cached for every split after that.
- 03
Render and check each stem
Hit Render preview to run the separation on your device. Use the synchronized stem mixer to solo, mute, and adjust the level of vocals, drums, bass, and other, and A/B against the original mix.
- 04
Export your stems
Download a single stem in any supported format, or grab all four as a ZIP of WAVs. You can also use Continue in another tool to send a stem straight to another tool without re-uploading it.
What you can do with this stem splitter
- Separate any track into 4 independent stems — vocals, drums, bass, and other — using the Demucs v4 Hybrid Transformer (htdemucs) model.
- Run the entire separation on your device via WebGPU, with an automatic WebAssembly fallback on devices that do not support it. Nothing is uploaded.
- Try a fast 30-second quick preview before committing to a full split, so you can judge quality without waiting for the whole song.
- Split tracks up to 10 minutes long, with a progress indicator that updates segment by segment as the model works through the file.
- Mix your split with a synchronized stem player: solo or mute any stem, set an individual level for each, and play them back together in sync.
- Export each stem separately in any of the supported output formats, or download all four at once as a ZIP archive of WAV files.
- Get sample-aligned, equal-length stems that line up exactly with the source track and with each other, ready to drop into a DAW.
- Send any stem straight to another tool with Continue in another tool, without exporting and re-uploading it yourself.
What is an AI stem splitter?
An AI stem splitter takes a single finished mix and pulls it back apart into the instruments that made it up. Instead of manually EQing around a vocal or hoping a mid-side trick isolates the bass, the tool uses a neural network trained to recognize what a vocal, a drum kit, a bass line, and everything else typically sound like inside a mix, then separates them into four independent audio files called stems.
The model behind this splitter is Demucs v4, specifically the Hybrid Transformer version (htdemucs). In plain terms, it looks at the audio two ways at once — as a raw waveform and as a frequency spectrum — then uses attention layers, the same basic idea behind modern language models, to work out which parts of the mix belong to which instrument. The whole process runs on your device: nothing is uploaded, and separation happens through your browser using WebGPU where it is available, falling back to WebAssembly (WASM) on devices that do not support it.
The 4 stems: vocals, drums, bass, and other
Every split produces exactly four stems, and each one is a full-length, standalone audio file you can play, mix, or export on its own.
| Stem | What it captures |
|---|---|
| Vocals | Lead and backing singing, spoken word, ad-libs — anything the model identifies as a voice. |
| Drums | Kick, snare, hi-hats, cymbals, toms, and other percussion. |
| Bass | The bass guitar or bass synth line, separated from the rest of the low end. |
| Other | Everything left over: guitars, piano, synths, strings, brass, pads — the full instrumental bed minus drums and bass. |
What people use stem splitting for
Splitting a track into stems opens up a few different workflows depending on what you are trying to do:
- Remixing — mute the original drums or bass and replace them with your own, keeping the vocal or instrumental bed intact.
- Sampling — pull a clean vocal phrase, a drum loop, or a bassline out of a track to chop up and reuse.
- Practice — mute an instrument to play or sing along with the rest of the band, or solo a part to learn it by ear.
- DJ edits and mashups — isolate the vocal from one track and the instrumental from another to build a blend.
- Transcription — solo the bass or drums to work out a part more accurately than you can from the full mix.
Quick preview, full splits, and the one-time model download
The first time you use the AI stem splitter, your browser downloads the separation model — about 172 MB — and caches it locally. Every split after that starts straight away, with no repeat download, as long as you use the same browser on the same device.
If you just want to hear whether a track will separate cleanly, turn on the quick preview option, which processes only the first 30 seconds. It is a fast way to judge the split before spending the time on a full render. Full splits are capped at 10 minutes per file, and progress is shown segment by segment while the model works, so you can see how far along a longer track is.
What to expect: bleed and separation quality
Source separation is genuinely good in 2026, but it is not perfect, and it helps to know what "good" looks like before you rely on it. The most common artifact is bleed — a trace of one instrument leaking into the wrong stem, most often a hint of vocal in the "other" stem or a faint cymbal in the bass. Dense mixes, heavily distorted instruments, and vocals that sit very close in frequency to a lead guitar or synth tend to separate less cleanly than a simply arranged track.
You will also hear this called splitting quality varying by genre and mix style. Sparse, well-produced pop and acoustic tracks usually split the cleanest. Wall-of-sound rock, dense electronic tracks, and anything with heavy vocal processing (doubling, wide reverb, pitch effects) are more likely to show audible bleed. Soloing each stem in the mixer before you export is the fastest way to catch this early.
Stem splitter vs. vocal remover: which one do you need?
These two tools use closely related technology but solve different problems. The stem splitter always gives you four separate stems — vocals, drums, bass, and other — which is what you want for remixing, sampling, or pulling out a specific instrument. The vocal remover is built for a narrower, two-output job: an instrumental (karaoke) track and an isolated vocal (acapella), nothing more. If all you need is a karaoke version of a song or a clean acapella, the vocal remover is faster and gets you there in one step; reach for the stem splitter when you need drums or bass on their own too.
Exporting stems for a DAW
Because every stem is sample-aligned with the source track and with each other, and all four are exactly the same length, you can drop them straight into a multitrack session and they will line up without any manual nudging. Export in WAV if you plan to edit or mix further, since it is uncompressed and keeps the alignment exact; use a compressed format like MP3 only for stems you intend to just listen to or share. For a full walkthrough of importing split stems into a session, see the guide to exporting stems to a DAW.
Tips for better stem separation
- Run the 30-second quick preview first. It is a fast way to judge whether a track will split cleanly before you commit to a full render.
- Solo each stem in the mixer after rendering and listen closely for bleed before you export — it is easier to catch problems early than after export.
- Only split audio you have the rights to use. Splitting a track does not change who owns the recording or the composition.
- For remixing, keep the "other" stem alongside drums and bass rather than discarding it — it usually carries most of the harmonic content of the track.
- On a phone or an older laptop, try the quick preview or a shorter clip first. Long files and low device memory can make a full split slow or cause it to run out of memory.
- Export as WAV if you are taking the stems into a DAW, and save a compressed format only for versions you plan to just share or listen back to.
Frequently asked questions
Is this AI stem splitter free?
Yes. Splitting a track into vocals, drums, bass, and other stems is free, with no account and no sign-up. The only one-time cost is bandwidth for the roughly 172 MB model download, which your browser caches after the first use.
Do my files get uploaded anywhere?
No. Separation runs entirely on your device using WebGPU or a WebAssembly fallback. Your audio file is never sent to a server — the model downloads once, and every split after that happens locally in your browser.
How long does splitting take?
It depends on the length of the track and your device. The 30-second quick preview is fast on most hardware. A full split of a longer file, up to the 10-minute limit, takes longer and is shown segment by segment so you can track progress as it works.
What is the difference between a stem splitter and a vocal remover?
The stem splitter always produces four stems — vocals, drums, bass, and other — for remixing or sampling any instrument. The vocal remover is built specifically for karaoke and acapella: it outputs just an instrumental and an isolated vocal, using the same underlying separation technology in a simpler, faster two-output mode.
How big is the AI model, and do I need to download it every time?
The Demucs v4 model is about 172 MB. It downloads once and is cached by your browser, so later splits start immediately. If you clear your browser cache or switch devices, the model downloads again on the next use.
What is the difference between the quick preview and a full split?
Quick preview processes only the first 30 seconds of a track, which is enough to judge separation quality quickly. A full split processes the whole file, up to a 10-minute limit, and produces complete, full-length stems.
Can I use the stems in my own music?
Yes, as long as you have the rights to the original recording — for example, a track you produced yourself, licensed audio, or something explicitly cleared for reuse. Splitting a song does not change who owns it, so only separate audio you have permission to use.
What file formats are supported for export?
Each stem can be exported in the formats this tool supports, including WAV (16-bit, 24-bit, or 32-bit float), MP3, FLAC, OGG Vorbis, Opus, M4A, and AIFF. You can also download all four stems at once as a ZIP of WAV files.
Is there a file size or length limit?
Files up to 150 MB are accepted, and a full stem split processes up to 10 minutes of audio per file. Use the quick preview option if you only need to check a shorter section, since it processes just the first 30 seconds.
Does this work on iPhone or Android?
It can, but AI separation is demanding, and mobile devices vary a lot in memory and processing power. Expect it to be slower than on a desktop, and try the quick preview first on a phone or tablet before attempting a full split of a longer file.
Why do my stems have some bleed between them?
A trace of one instrument leaking into another stem — a hint of vocal in the "other" stem, for example — is a normal limitation of current AI separation, not a bug. Dense mixes and instruments that sit close in frequency to the vocal tend to show more bleed than simply arranged tracks.
Can I download all four stems at once?
Yes. Instead of exporting each stem individually, you can download all four — vocals, drums, bass, and other — together as a single ZIP archive of WAV files, which keeps them aligned, organized, and ready to drop into a session.
My browser ran out of memory during a split. What should I do?
Try the 30-second quick preview instead of a full split, close other browser tabs and apps to free up memory, or use a shorter clip of the track. Long files combined with limited device memory are the most common cause of this, especially on phones and older laptops.
Can I send a stem to another tool for further editing?
Yes. Use Continue in another tool on any rendered stem to hand it off directly to another tool, such as the pitch shifter or tempo changer, without exporting the file and re-uploading it yourself. The handoff happens locally, stored in your browser.