AI vocal remover

The same separation models the paid services run, without the credit system on top.

Drop your song here

or click to choose · mp3, wav, flac, m4a · up to 50 MB

What do you want?

VocalsTake the singing out, or keep only the singing.You get 2 files:VocalsSong without vocalsDrumsGet the drums alone, or play along without them.You get 2 files:DrumsSong without drumsBassGet the bassline alone, or play along without it.You get 2 files:BassSong without bassEverythingEvery part on its own, and the song minus each part.You get 8 files:VocalsDrumsBassOther (guitars, keys, synths)Song without vocalsSong without drumsSong without bassSong without the rest

Almost every vocal remover you can find today is running one of a small number of open machine-learning models. The differences between the services are rarely the technology — they are the pricing page, the queue, and how much quality is withheld until you subscribe.

This one runs those models and gives you the full-quality output. That is the entire proposition, and the rest of this page explains what the AI is actually doing so you can judge the results rather than take a claim on faith.

What the model is actually doing

A separation model does not filter frequencies or invert phase — the tricks that older vocal removers relied on, and the reason those only ever worked on a narrow set of recordings. It has been trained on a very large number of songs for which both the finished mix and the original separate tracks were available, and from that it learned what a human voice looks like inside a spectrogram.

Given a new mix it has never encountered, it predicts which parts of the sound belong to the voice and which belong to everything else, then reconstructs two separate recordings from that prediction. Because it is a prediction rather than an unmixing, the output is an estimate — which is why quality varies with how closely a song resembles what the model was trained on.

Two models do the work here. A specialist MDX-Net model handles the two-way voice-and-backing split, because it is both faster and cleaner at that specific job. Meta's Demucs handles the four-way split into vocals, drums, bass and everything else.

Using it

1
Upload. Any common audio format, up to 50 MB.
2
Pick what you want. Vocals is the fast two-file split; the other options run the four-part model.
3
Confirm your email. One six-digit code, once per browser.
4
Download. 320 kbps mp3, with nothing withheld for a paid tier.

Questions

›How is this free when everyone else charges?

The models are open source, so the technology costs nothing to use. What the paid services are really billing for is GPU time. That cost is real but small — well under a cent for a typical song — and it is covered here rather than passed on.

›Is the quality actually the same as the paid tools?

For the models in use, yes — there is no reduced-quality free tier. Some paid services do offer newer or fine-tuned models that can edge ahead on difficult material; on ordinary songs the difference is hard to hear.

›Does it get better over time?

The models improve as the research does, not as you use them. Nothing you upload is used for training, and no personal model is built from your files.

›Is my audio used to train anything?

No. Your upload is deleted as soon as it has been processed, and the separated files go once your download links expire.

›Why is there an email step at all?

It is how the daily limit is tracked without charging anyone, and it is the only thing asked for. No password, no card, no trial that expires into a bill.

Where it struggles

No model is magic, and any service claiming otherwise is overselling. Separation quality tracks how conventional the production is: a modern studio mix separates beautifully, a bootleg live recording does not. Expect artefacts on heavily processed vocals, on very quiet passages, and anywhere the voice and an instrument occupy the same narrow band of the spectrum.