AI Vocal Remover & Karaoke Maker

Split a song into an instrumental and an acapella with a source separation model running in your browser. WAV output, batch friendly, nothing uploaded.

🌐 Español

Drop your files here (.mp3, .wav, .m4a, .ogg, .flac)

🔒 Private by design: your files are processed locally in your browser and never uploaded to any server.

Written and maintained by Last updated

Separating a mix back into its parts

Once a song is mixed down to two channels, the individual instruments are gone as separate entities. What remains is the sum of everything, and pulling one element back out of a sum is not an operation with an exact answer.

For years the only available trick was channel cancellation, which relies on vocals usually being panned to the centre. Subtract one channel from the other and anything identical in both disappears. It works, in the sense that the voice gets quieter, and it also removes the kick drum, the bass and the snare, which are centred for the same reason, and it leaves you with a mono file.

A trained separation model works differently. It has learned, from a large number of songs where both the mix and the separate parts were available, what a human voice looks like as a pattern of energy across frequency and time. Given a new mix it predicts which parts of that pattern belong to the voice, and the prediction is made per frequency and per moment rather than per channel. The drums stay, the bass stays, and the result stays in stereo.

Separating a file

  1. Drop in one or more MP3, WAV, M4A, OGG or FLAC files.
  2. Choose What to extract, an instrumental, the vocals, or both.
  3. Run it. The first run downloads the model, and separation itself takes real time.
  4. Download the WAV, or the zip when you asked for both.

What the model is actually doing

The audio is first decoded to stereo at 44.1 kHz, which is the rate the model expects. Mono material is duplicated into both channels and anything with more than two channels is reduced to two.

That waveform is then converted into a spectrogram, a picture of how much energy sits at each frequency at each instant. This is the representation the model works in, and it is why the approach can separate things that overlap in time: two sounds happening simultaneously usually occupy different frequencies, and where they overlap the model has learned which is more likely to be which.

The model outputs a mask, a value per frequency per frame saying what proportion belongs to the voice. Applying the mask and converting back to a waveform produces the separated audio. The other part is what remains.

Why the output is uncompressed

The separated audio arrives as raw samples and is written directly to a WAV.

Encoding it to MP3 at that point would layer compression artefacts on top of separation artefacts, and the two do not hide each other. Separation leaves faint spectral residue in unusual places, precisely the kind of thing a lossy encoder handles badly, and the combination sounds noticeably worse than either alone.

The files are large as a result. If you want them smaller, convert afterwards as a deliberate step with MP3 to WAV in reverse, once you have confirmed the separation is good enough to keep.

What to expect, honestly

The instrumental will have faint traces of the voice in it, mostly in loud sections and on reverb tails, because reverb spreads a voice across time and frequency in ways that are hard to attribute. The acapella will carry a little of the backing for the same reason.

Sparse arrangements separate better than dense ones. An acoustic recording with a clear vocal comes out very cleanly. A modern production with heavy compression, layered harmonies and effects on everything is a much harder problem.

None of this is a limitation of this particular tool so much as of the task. Reconstructing parts from a mix is inference, and no amount of processing recovers information that the mixdown discarded.

If the source is quiet or uneven before you start, Normalize Audio gives the model a better input to work from, and Remove Audio Noise helps with a noisy recording. If one channel of your source is dead, fix that first with Convert Stereo to Mono, since a broken stereo image confuses the separation.

To get the words as text rather than as audio, Transcribe Audio runs a different model in the same local way, and works noticeably better on a separated vocal than on a full mix. The rest are on the audio hub.

Frequently asked questions

How is this different from the old trick of cancelling the channels?

Entirely different, and much better. The old method subtracts one stereo channel from the other, which removes anything sitting exactly in the centre. That takes out the vocal and also the kick drum, the bass and the snare, and it collapses the result to mono. This uses a trained model that separates by learning what a voice looks like in the frequency domain, so the rhythm section survives and the result stays stereo.

What actually happens to the audio?

It is decoded to stereo at 44.1 kHz, converted into a time and frequency representation, and the model predicts a mask over that representation saying how much of each frequency at each moment belongs to the voice. The mask is applied and the result converted back to a waveform. Mono input is duplicated across both channels first, and anything with more than two channels is reduced to two.

Why is the output a WAV rather than an MP3?

Because the separated audio comes out of the model as raw samples and writing them straight to a WAV avoids adding another lossy encode on top of whatever the source already suffered. Separation artefacts and MP3 artefacts compound unpleasantly. The files are large, and converting them afterwards is a deliberate second step rather than something forced on you.

What do I get when I ask for both?

A single zip containing the instrumental and the vocal, both named after your original file. Asking for one gives you that one as a plain WAV with no zip around it. The choice only affects packaging, since the separation itself always produces both parts internally.

How good is the result?

Good enough to sing over, and audibly imperfect if you listen closely. Expect faint traces of the voice left in the instrumental, especially in loud choruses and on reverb tails, and expect the acapella to carry a little of the backing. Dense, heavily processed modern productions are harder than sparse ones. It will not match a studio master built from the original multitrack, because that information is not recoverable from a mixed file.

Is my music uploaded?

No. Two model files of roughly 20 MB each are fetched once from a public model host and everything after that runs on your own device. The audio never leaves the machine, which matters for anyone working on their own unreleased recordings rather than on commercial tracks.

How long does it take?

Longer than you expect, because separation is genuinely heavy work and it runs on a single thread inside a browser tab. A full length song takes a real amount of time rather than seconds. The tool shows progress across the batch and can be cancelled. Running several files means running them one after another on your own processor rather than queuing them on a server.

Related tools