How to use this workflow
Choose a local recording
Select a supported local file up to 250 MB. Device memory and duration checks may reject a file below that size. Begin with a short excerpt.
Start separation and allow the model to load
Select Start separation. The tool checks or downloads an approximately 67 MB model plus runtime files. It prefers a compatible GPU and offers a CPU/WASM compatibility path when available. Keep the page open.
Listen to both outputs
Switch between the instrumental and vocal previews. Check quiet and loud sections, then download the desired WAV or MP3 output. To compare with the source, play your original file separately; this result player has two output choices.
What the model produces
The model estimates the instrumental component. The vocal output is derived from the difference between the source and that estimate, so a sound missing from the instrumental can appear in the vocal result even if it is not a voice.
Audio is prepared as a 44.1 kHz stereo model input. The process is not equivalent to obtaining the original studio vocal and instrumental tracks, and it does not identify or license a song for you.
Check the result before using it
Backing vocals, doubled voices and long vocal reverbs can remain audible in the instrumental. Some instruments may be weakened or sound watery after separation. Listen to the intended output in context rather than judging only a silent gap.
The audio computation takes place on the device. Model and runtime downloads still require network access, and the browser may cache model files or use temporary storage. Long recordings can require substantial memory; a compatible browser is not a promise that every phone will complete every file.
Common questions
Can I remove every trace of a singer?
There is no guarantee of complete removal. Vocal effects and overlapping instruments can leave residual sound or cause missing musical detail.
Should I use the four-stem splitter instead?
Use it when you also need separate drums, bass or other instruments. The four-stem tool requires WebGPU and has its own model and limits.