Audio Splitter

AI Source Separation (experimental, opt-in)

The default model is HTDemucs v4 (6-stem) — 246 MB, hosted on Hugging Face by MansfieldPlumbing/Demucs_v4_TRT. Click Load model to download it once; it's then cached in the browser's Cache API so subsequent visits are instant. Splits each track into drums, bass, other, vocals, guitar, piano. Long tracks are processed in 7.8-second overlapping chunks with crossfade. Everything still runs entirely on your device — no audio is uploaded.

How it works. This page runs entirely in your browser using the Web Audio API and pure-JS DSP — your audio is never uploaded. Stems are derived with classical signal-processing tricks: Instrumental uses the L−R cancellation technique (removes any centered mono content, which on most pop tracks is the lead vocal). Vocals approximates the inverse: it keeps the centered mono component (mid) and subtracts the side channel. Bass is a 200 Hz lowpass of the full mix; Highs is a 6 kHz highpass. Quality varies wildly by track — well-mixed pop with a centered, dry vocal works best; live recordings, hard-panned vocals, and heavy reverb will defeat the centre-cancel trick. For true source separation (Demucs / Open-Unmix quality) flip the Mode switch to AI at the top of the page. That lazily loads onnxruntime-web and fetches an ONNX model URL you provide; the model is cached in the browser so subsequent visits are instant. Nothing is downloaded unless you opt in.