Free acapella extractor

What is an acapella extractor?

An acapella is a vocal on its own, with no music under it. An acapella extractor, also called a vocal isolator, vocal extractor or voice isolator, takes a finished song and estimates that vocal for you with a neural network. To extract vocals from a song, you drop the stereo file and you get the voice back as a separate file you can work with.

This tool is made for songs. If you want to clean the noise out of a podcast or a voice memo, a speech-enhancement tool is a better fit. It can pull a voice out of a video, but it is trained on music, not on rooms and microphones.

What to do with an acapella

Remix and edit: put the vocal over a new beat, change its tempo and key in your DAW and build a different song under it. Mashup: pair an acapella with the instrumental of another track that shares its key and tempo. Chops and samples: cut a phrase, a word or a breath and play it from a pad. Study: hear how a vocal was tuned, doubled and treated, with nothing to cover it.

You can also isolate the voice of your own demo to rework the beat under it, or to send a singer a guide track without the old arrangement.

Getting a cleaner acapella

The cleanest results start with the best source. Use a WAV or a high-bitrate file instead of a low-bitrate MP3, and pick songs with a vocal that sits clearly in front. A voice buried in loud guitars, or a heavily layered chorus, will leave more traces than a verse with a sparse backing.

Listen to the acapella alone, then switch to the original and back, so you hear what the model kept. If the instrumental comes through faintly, a gentle gate between phrases, a low cut and a de-reverb pass in your DAW will tidy it up. If you also want to check that nothing is missing, play the vocal with the instrumental from the vocal remover page: together they should sound like the song. For the cleanest vocal pick HQ: on our test song with sung vocals it left silence between the phrases where Fast left traces of the music.

How the vocal gets isolated

Fast uses Demucs v4, a hybrid transformer from Meta's AI research team. It reads the song as a waveform and as a spectrogram, and it was trained on multitrack recordings, so it learned how voices differ from drums, bass and other instruments. HQ uses BS-RoFormer, a newer band-split transformer that scores higher on public benchmarks. For each stretch of the song they predict the vocal, and the pieces are cross-faded into one smooth file.

It runs on your device. With a computer that supports WebGPU it uses the graphics card, which HQ needs, and Fast can also use the processor. It treats lead and backing vocals alike, so you get one vocal file with every voice in it. Double-tracks, harmonies and ad-libs stay together.

About rights

An acapella of someone else's song is still their work. Use this on music you own or have permission to use, and clear any release or public remix with the rights holders. Many artists and labels publish official acapellas or stems for remix contests, and those are the safest to build on.

Private, free and in beta

The vocal is isolated on your device, so the song is never uploaded and we can keep the tool free, with no sign-up, no credit counter and no daily cap. The limit is 10 minutes per song, or 6 in HQ. It is a beta: it works, it is new and it is improving. Write to info@obra.art if you have a song it handles badly.