Remove vocals from a song

Remove the vocals from any song in your browser and keep the instrumental, or keep only the vocals. An AI model separates them on your own computer. Free, no upload.

AudioRinse · vocal removerOn-device · 0 B uploaded

Drop a song here

MP3 · WAV · M4A · MP4 · MOV · WEBM — up to 15 min. Nothing is uploaded.

The model downloads once and stays in this browser, so the next song starts at once. Your song is never uploaded.

How it works

  1. Drop a song onto the box above: MP3, WAV, M4A, FLAC, or a video.
  2. An AI model splits it into vocals and instrumental on your computer. The first time, the model downloads once.
  3. Listen, switching the vocals and the instrumental on and off. Voice still showing through? Separate again in Best quality.
  4. Download the instrumental, the vocals, or both in a ZIP, as MP3, M4A or WAV.

How it takes the vocals out

AudioRinse runs a source-separation model in your browser: a neural network trained on thousands of songs to tell a voice from the instruments around it. It hears the singer wherever they sit in the mix, with their reverb and harmonies, and writes the vocals to one track. The instrumental is the song minus those vocals, so the two always add back up to the original, sample for sample.

Old vocal removers flipped one stereo channel against the other to cancel whatever sits in the middle. That takes the bass and the kick drum with the voice, and leaves any vocal recorded in stereo. A model doesn't depend on where the voice is panned, so the instrumental keeps its low end.

Standard or Best

  • Standard runs HTDemucs, Meta's open separation model (126 MB). It works on any computer, on the graphics chip through WebGPU where the browser has it and on the processor where it doesn't.
  • Best runs Mel-Band RoFormer, a newer model (464 MB). On the MUSDB18 benchmark songs its vocals measured about 3 dB cleaner than Standard's and its instrumental about 2.7 dB cleaner: less of the voice left behind, fewer cymbals in the vocals. It takes about three times as long and needs a fast graphics chip, so the page offers it only where yours is.
  • Try Standard first. If you hear the voice in the instrumental, the result screen offers to separate the song again in Best.

Getting a clean instrumental

  • Start from the best copy you have: a WAV, FLAC or high-bitrate MP3 separates more cleanly than a low-bitrate rip, whose artifacts the model can't tell from the voice.
  • Reverb, delay and backing vocals belong to the voice and go with it. A long, loud reverb tail can leave a faint trace.
  • An instrument that doubles the melody, like a synth lead or a sax playing along with the singer, may stay in the instrumental or come out partly in the vocals.

What it's for

Karaoke and backing tracks, practising your instrument with the band, covers, DJ edits and remixes, or the acapella for a mashup. Separating a song doesn't change who owns it: use the stems for your own practice, or with the rights holder's permission when you publish.

Questions

Is my song uploaded anywhere?

No. The separation model runs inside your browser, on your own computer. Your song never leaves your device, so there's no upload to wait for and nothing stored on a server. The only download is the model itself, the first time.

How clean is the instrumental?

Clean on most songs, with faint traces of the voice where it's loudest or drenched in reverb. Measured on the MUSDB18 test songs, the usual benchmark, Standard's instrumental scores 13.9 dB and Best's 16.6 dB (signal-to-distortion ratio, higher is cleaner). Listen before you download: switch the vocals off and on to compare.

Can I keep only the vocals?

Yes. Switch the instrumental off and the vocals on, and the download is the vocals alone: an acapella. Or use the acapella extractor, which starts there.

Can I also split out the drums and bass?

Yes, with the stem splitter: it separates a song into vocals, drums, bass and everything else.

Does it work on a video?

Yes. Drop the video: its sound is separated, and you can download the video with only the instrumental (or only the vocals) as its sound, the picture copied as it is.

How fast is it?

It depends on your graphics chip. A 4-minute song takes about 50 seconds on a gaming laptop's graphics card and about 3½ minutes on a 2020 laptop's built-in graphics; Best quality takes about three times as long where it's offered. Without WebGPU (some browsers, older computers) it runs on the processor: about 7 minutes on an 8-core laptop. The page shows the time left as it goes.

Why does the first song take longer?

The model downloads once: 126 MB for Standard, 464 MB for Best. It stays in your browser, so the next song starts at once, and it works with Wi-Fi off from then on.

How long can the song be?

Up to 15 min. Free downloads cover songs up to 10 minutes. On a computer with little memory the page takes less, and says so when you drop a longer file.

Is it free?

Separating and listening are free and unlimited. You get 3 free downloads a month for songs up to 10 minutes (1 without an account), and all the stems of one song count as one download. Pro adds unlimited downloads, songs up to 15 min and video.

Other tools