Split any song into an isolated acapella and a clean instrumental with the htdemucs AI model - running on YOUR graphics card, so the song never uploads anywhere.
100% in your browser - your files never leave your device.
Any MP3, WAV, OGG, FLAC or M4A. It is decoded locally; nothing is uploaded, which is why there are no queues, accounts or per-song limits.
The first run downloads the open htdemucs model (~120 MB, cached for next time). Inference runs through WebGPU on your graphics card - a modern GPU separates faster than the song plays.
You get two studio-ready WAVs: the isolated vocals and everything else summed into an instrumental. Preview both in the page before downloading.
Karaoke and live covers are the classic use: strip the lead vocal from the original and sing over a full-quality instrumental instead of a MIDI backing. Producers go the other way - the isolated acapella becomes remix and mashup material, ready to drop over a new beat (check its key and tempo with the BPM & key finder first).
Practice is the quieter superpower: mute the part you are learning and play along with the rest of the band. Singers rehearse against the instrumental; transcribers isolate the vocal line to hear phrasing and harmonies that are buried in the mix.
Modern separation is a neural network trained on multitracks, not the old phase-inversion trick that only worked on center-panned vocals and gutted the bass. htdemucs (Demucs v4, the open model this tool runs) reconstructs each part, so reverb tails and doubled vocals come out with far fewer artifacts. And because everything runs in your browser, unreleased demos and client work stay on your machine.
No. The AI model is downloaded TO your browser once (~120 MB, cached) and inference runs on your graphics card via WebGPU. The audio never leaves your device - there is nothing to queue, pay for or delete afterwards.
The neural network is far too heavy for plain JavaScript, so it runs on the GPU through WebGPU. Up-to-date desktop Chrome and Edge support it out of the box (Safari 18+ too); the page will tell you if your browser cannot run it.
The tool runs htdemucs (Demucs v4), one of the strongest open-source separation models. On typical modern mixes the instrumental is clean enough for karaoke and remixing; very dense or heavily-mastered material can leave slight vocal residue - true of every separator.
After the one-time model download, separation speed depends on your graphics card - a recent discrete GPU processes faster than realtime, integrated graphics take a few times the song length. The progress bar tracks it chunk by chunk.
Yes - the same model actually produces four parts. Use the stem splitter page to download drums, bass, other instruments and vocals as individual WAVs.
Open the stems in TuttiDAW and you have a full remix session - EQ, effects, arrangement - without leaving the browser.
Explore TuttiDAW - the free browser DAW