Acapella extractor: isolate the vocals from any song in your browser

There is no vocal track hidden inside an MP3 waiting to be found; the voice was mixed together with everything else. SplitVocals runs Demucs v4, a hybrid-transformer model trained on thousands of songs to recognise what a singing voice sounds like inside a full arrangement, and reconstructs it as a separate stream. The lead vocal, the harmonies and the ad-libs all count as voice and come out together in the vocal stem; drums, bass and the rest of the band land in their own stems.

What you get is a standard audio file — WAV at the source sample rate or 320k MP3 — that drops straight into a DAW, a sampler or a DJ deck. The other three stems come from the same pass, so the instrumental is one click away if you need it too.

Producers and remixers lay a studio vocal over a new beat, retime it, or chop it into a hook. DJs blend an acapella over another record's instrumental for a live mashup. Singers and vocal coaches study phrasing, breath and riffs with the band out of the way, then sing along to the isolated line. Songwriters and transcribers hear a lyric or a harmony part that the mix buried. Sound designers and video editors pull a spoken or sung line to use as a texture. None of these needs an account; the separation runs in the browser tab.

A clean acapella starts with a clean recording. Studio vocals that were tracked on their own microphone separate almost completely; live recordings, where the singer bleeds into every other mic on stage, do not. Reverb and delay on the voice are part of the voice — the model keeps the effect with the vocal, and there is no way to get a drier take than the one that was recorded. Instruments that live in the same range as the voice, such as a lead guitar or a synth lead doubling the melody, can leave faint traces in the stem.

Give the model the best source you have: a lossless WAV or FLAC, or a streaming-quality file rather than a low-bitrate rip. If a vocal still comes back with bleed, the HD Pack and Pro run a larger model on a server that separates vocals noticeably more cleanly, and Pro adds a six-stem split that also isolates guitar and piano.

Producers often need an acapella from a demo, a session bounce or a vocal that has not been released. On this page the file is decoded and processed inside your browser through WebGPU or WebAssembly and is never uploaded — there is no server copy to leak, expire or sit in someone's queue. Open the Network tab of your developer tools during a separation and you will see no outgoing request with your audio in it.

It is also faster: no upload of a large WAV, and once the ~170 MB model is cached after the first run, every separation starts instantly, even offline. A typical song takes one to three minutes on Chrome or Edge with WebGPU; Safari and Firefox use a slower CPU path.