Separate a Song Into Stems with AI

Stem separation normally means uploading your music to somebody else's server and paying per track. This runs the model on your own machine instead. Take the vocals out to get a karaoke backing track, pull the vocal up on its own for a remix or a sample, or split the whole song into drums, bass, vocals and everything else. The model is downloaded once, cached, and works offline afterwards — and because nothing is uploaded, unreleased material and client work can go through it too.

How to Separate a Song Into Stems with AI

  1. 1

    Choose what you want back

    Two tracks (vocals and everything else) is the karaoke split and uses about half the memory. Four tracks gives you drums, bass, vocals and other separately.

  2. 2

    Download the model

    The model is around 166 MB, fetched once and cached in your browser. Later separations start immediately and work with no connection.

  3. 3

    Add your songs

    Drop in one file or a whole batch, each up to six minutes. MP3, WAV, M4A, OGG and FLAC all work, mixed freely.

  4. 4

    Separate and download

    Press Separate and leave it. The queue runs on its own, the tab title carries the progress so you can work in another window, and you can keep adding files while it goes. Download each track as MP3 or WAV, or grab everything at once.

How it works

This runs HT-Demucs, the hybrid transformer version of Demucs from Meta's research team, compiled from C++ to WebAssembly and executed in a Web Worker. It is a genuine AI-and-DSP hybrid rather than a pure neural network: the model works in both the spectrogram domain and directly on the waveform, combining the two, which is what lets it keep transients like drum hits sharp instead of smearing them the way a spectrogram-only approach does. Classical signal processing does the surrounding work — your file is decoded and resampled to the 44.1 kHz stereo the model expects, processed in overlapping windows, and the pieces are cross-faded back together so the joins are inaudible. Separation is genuinely heavy: it runs at roughly three times slower than real time, so a three-minute song takes something like nine minutes. That is the trade for running a model of this size on your own CPU instead of a rented GPU, and the tool tells you the estimate for your specific file before you start.

Supported formats

WAV, MP3, M4A, AAC, OGG and FLAC. Files are decoded by your browser, so anything it can play will work here. Output is exported as WAV.

Frequently asked questions

Is my audio uploaded anywhere?
No. Everything runs inside your browser using the Web Audio API, so the file never leaves your device. You can prove it: load the page, turn off your Wi-Fi, and the tool still works.
How do I make a karaoke track?
Choose the two-track mode. You get the instrumental — the song with the voice removed — plus the isolated vocal as a separate file. The instrumental is the karaoke backing track.
Why is there a six-minute limit?
It is a memory limit rather than a policy. Separation holds the entire song in memory as several uncompressed tracks at once, and browsers run out of room before a long song finishes. The two-track mode uses about half as much, so it copes with more.
Why does it take so long?
Separation runs at roughly three times slower than real time, so a three-minute song takes around nine minutes. The model is doing the same work a paid service does, except on your CPU rather than a rented GPU, and the engine is built to fit in a browser's memory rather than to be fast. The trade is time for privacy and cost: your music is never uploaded and there is no per-track fee. The tool shows an estimate for your file before you commit to it.
How good is the separation?
Demucs is among the strongest openly available models for this, and drums and bass in particular come out very cleanly. Isolated vocals can carry some bleed from instruments that occupy the same register — brass and accordion are the usual culprits — which is normal for every separation tool, paid ones included.
Can I separate a whole batch at once?
Yes, and given how long each song takes it is the sensible way to use this. Queue up as many files as you like and start it — they are processed one after another, the tab title shows progress so you can leave it in the background, and you can keep adding files while it works. If one file fails or is too long it is skipped and the rest carry on, so a bad file cannot waste an unattended run.
Does this use AI?
Yes, and it is the most demanding AI in the app. HT-Demucs is a transformer-based model that works in both the spectrogram and waveform domains at once, which is why it keeps drum transients crisp. Traditional DSP still does the surrounding work: resampling, segmenting, and the overlapping cross-fade that stitches the separated segments back together.
Do I need an account?
No. The free tools work straight away with no sign-up. Creating a free account raises your daily processing limit from 20 to 60 minutes.
Is it free?
Yes. Every tool on this page is free to use, with a daily limit on total audio processed. There is no watermark and no trial period.

Other free audio tools