Is it really free? Do I need to sign up?

It’s completely free with no account needed. Just open the page and pick a file.

Is the song I choose sent anywhere?

No. All separation happens entirely inside your browser.

Which file formats are supported?

MP3, WAV, M4A, AAC, FLAC and OGG, plus the audio track of video files.

Which model should I pick?

If you are unsure, take Standard — the one marked Recommended. Pick Fast when speed matters most, 4stem or 6stem to split out drums, bass and the rest, or High Quality when you want the best result and can wait. Every card shows what comes back and roughly how long it takes.

How long does it take?

It depends on the model and your device. Fast handles a 3-minute song in well under a minute; Standard takes a few minutes. Each card shows an estimate for your device before you start, and you can keep working in other tabs meanwhile.

Can I use it on a phone?

Yes. On iPhone and Android, Fast splits vocals from the backing track. Models that need more memory than your device has are shown as unavailable, with the reason.

The separated parts sound slightly different from the record.

Separation does not take a mix apart; it estimates how each part was sounding. Being an estimate, it never matches the recording exactly. The moments it finds hardest are where several instruments play the same pitch at the same time, and that is where the difference shows. Heavier models — High Quality, 6stem — narrow the gap without closing it.

It stops partway through.

Almost always device memory. Closing other tabs often gets it through. If not, drop one step lighter — 6stem to 4stem, Standard to Fast. Long tracks cost the most, so trimming to the section you actually need with the audio trimmer first is another way through.

What format do I get, and can I have MP3?

WAV, 16-bit stereo. Nothing is compressed, so it loads straight into a DAW or a sampler. For something smaller, or for your phone, the file converter on this site turns it into MP3 — also entirely in the browser, with no upload.

Can I publish or distribute what I separate?

Separating a song does not change who owns it. Publishing, distributing or using the result commercially is yours to clear. Please use the tool on audio you own, audio you have permission to use, or within the scope of private use. The Terms of Use set this out in full.

Does it work offline?

The processing itself runs on your device and never uses the network, and the model stays on your device after the first download. Starting a separation, however, always loads the engine runtime files (tens of MB) from the network, and those are not stored for offline use — so a separation cannot be started with no connection. The tools under Kit are the ones that keep working offline once you have used them.

Runs in your browser

Stem & instrument — separation, in your browser.

Input

Drag & drop an audio file

or click to choose

MP3, WAV, M4A, AAC, FLAC, OGG, Opus, AIFF and more — plus the audio of MP4, MOV, AVI, WebM, MKV and other video

Six ways to split a track

FastFastest. Works on phones tooVocals + backing
4stem FastFour parts, still light — works on phonesVocals, drums, bass, other
StandardPick this if unsure — quality without the waitVocals + backing
4stemSplit into band partsVocals, drums, bass, other
6stemDown to guitar and piano6 parts
High QualityBest possible — and it takes a whileVocals + backing

Drop in a track and you pick from these. Estimated run times appear then, along with a reason for anything this device does not have the memory to run.

Result

Your separated parts will appear hereChoose a file and a model, and vocals, drums, bass and the rest come back as separate tracks. Mix them, and export only the ones you need.

Vocals and backing, of course — but also drums, bass, guitar and piano. Up to 6 parts, and you export only the ones you need.

What you can do

Up to six parts: Not just vocals and backing — pull out drums, bass, guitar and piano on their own.
No sign-up needed: No account, no install. Just open the page and choose a file.
For karaoke and transcribing: Remove vocals to make a backing track, or isolate just the vocals. Great prep for practice or production.
Preview and adjust before export: Try each part’s volume and mute right there, then save only what you need as WAV.

What is 6-stem?

It means automatically splitting one song into six parts. Here is what each one is:

Vocals: The sung/lead voice part.
Drums: Percussion such as kick, snare and hi-hats.
Bass: The low-end string/bass part.
Guitar: Electric and acoustic guitar.
Piano: Keyboard/piano parts.
Other: Backing elements that don’t fall into the categories above.

How to use

Choose a file
Drag & drop an MP3, WAV, M4A, FLAC or video, or click to select.
Pick a model
Not sure? Take Standard, the recommended one. In a hurry, take Fast; to split by instrument, take 4stem or 6stem.
Listen and export
Play and adjust each part, then save just what you need as WAV.

How separation works

What is actually happening when a tool "removes the vocal", described as it is implemented.

It estimates rather than subtracts: The old method subtracted one channel from the other to cancel whatever sat in the centre. That does remove a centred vocal — along with the kick, the snare and the bass, which are centred too. What runs here is different: a trained model estimates, for every moment and every frequency, how much of that energy belongs to each part, and divides the audio accordingly. That is why the backing keeps its weight.
The song is opened into a time-frequency grid: A short-time Fourier transform (STFT) turns audio into a grid of when, at what pitch, and how much. The model reads that grid and returns each part’s share of it — a mask. Multiplying the mask into the original grid and transforming back gives that part on its own. Spleeter, which Fast uses, works this way. htdemucs, used from Standard upward, keeps the transform inside the model: waveform in, waveform out.
It runs on your device, a slice at a time: The models are converted to ONNX and executed by ONNX Runtime Web — on the GPU through WebGPU where the browser supports it, on the CPU otherwise. The work happens in a worker separate from the page, and the song is cut into short segments that go through transform, estimate and reconstruct in turn. Nothing expands the whole song at once: doing that needs close to a gigabyte of working memory for a three-minute track, and it used to fail there.
Adding the stems back does not restore the original: Because separation is estimation, summing the parts does not reproduce the source file. It sounds like the same song; as a waveform it carries the estimation error. Where you need the original audio, use the original file rather than a reconstruction.
A model is fetched once: Model files download to your browser directly from where they are published (Hugging Face), not from our servers. Each is roughly 20–260 MB, and once fetched it stays in the browser cache, so later runs start immediately. You can see how much is stored, and delete it, from the privacy policy page.

Frequently asked questions

Is it really free? Do I need to sign up?: It’s completely free with no account needed. Just open the page and pick a file.
Is the song I choose sent anywhere?: No. All separation happens entirely inside your browser.
Which file formats are supported?: MP3, WAV, M4A, AAC, FLAC and OGG, plus the audio track of video files.
Which model should I pick?: If you are unsure, take Standard — the one marked Recommended. Pick Fast when speed matters most, 4stem or 6stem to split out drums, bass and the rest, or High Quality when you want the best result and can wait. Every card shows what comes back and roughly how long it takes.
How long does it take?: It depends on the model and your device. Fast handles a 3-minute song in well under a minute; Standard takes a few minutes. Each card shows an estimate for your device before you start, and you can keep working in other tabs meanwhile.
Can I use it on a phone?: Yes. On iPhone and Android, Fast splits vocals from the backing track. Models that need more memory than your device has are shown as unavailable, with the reason.
The separated parts sound slightly different from the record.: Separation does not take a mix apart; it estimates how each part was sounding. Being an estimate, it never matches the recording exactly. The moments it finds hardest are where several instruments play the same pitch at the same time, and that is where the difference shows. Heavier models — High Quality, 6stem — narrow the gap without closing it.
It stops partway through.: Almost always device memory. Closing other tabs often gets it through. If not, drop one step lighter — 6stem to 4stem, Standard to Fast. Long tracks cost the most, so trimming to the section you actually need with the audio trimmer first is another way through.
What format do I get, and can I have MP3?: WAV, 16-bit stereo. Nothing is compressed, so it loads straight into a DAW or a sampler. For something smaller, or for your phone, the file converter on this site turns it into MP3 — also entirely in the browser, with no upload.
Can I publish or distribute what I separate?: Separating a song does not change who owns it. Publishing, distributing or using the result commercially is yours to clear. Please use the tool on audio you own, audio you have permission to use, or within the scope of private use. The Terms of Use set this out in full.
Does it work offline?: The processing itself runs on your device and never uses the network, and the model stays on your device after the first download. Starting a separation, however, always loads the engine runtime files (tens of MB) from the network, and those are not stored for offline use — so a separation cannot be started with no connection. The tools under Kit are the ones that keep working offline once you have used them.

Please only use audio you own the rights to, or that you are otherwise permitted to use.