What you get
Preview and download two WAV stems: estimated vocals and the residual accompaniment. Results depend on the mix and may contain artifacts.
Local processing
The tool uses the Open-Unmix UMX-HQ vocals network exported to ONNX and ONNX Runtime Web. The model is fetched only when you press Separate; no paid inference API or subscription is used. Audio is decoded, processed and exported locally.
Questions
Is my audio uploaded?
No. The selected audio is decoded and processed in your browser. The only network request is the model download when you start separation.
Why is the first run slow?
The browser downloads a 35.6 MB model once; the browser may cache it. Inference then runs locally on the device CPU.
Will every song separate cleanly?
No. The model estimates a vocal stem and may leave vocal leakage, remove instrument harmonics or create artifacts. Listen before relying on a stem.
Scope and limitations
This is an estimate for stereo 44.1 kHz processing and clips up to 30 seconds. Mono sources are duplicated to stereo. Vocals mixed deeply into instruments, heavy effects, live recordings and unusual genres can produce leakage or artifacts. Do not use the result as a clean master. Model download requires an internet connection; audio does not leave this device.