Vocals
100%
Local AI audio tool
Four-Stem Splitter
Separate one song into vocals, drums, bass, and other. Your audio stays on this device from upload to download.
VocalsDrumsBassOther
Desktop browser required
This tool needs desktop Chrome or Edge
Four-stem processing uses browser features that are not available in this browser.
No model or audio has been downloaded or transmitted.
Four separated stems will appear here
00:00.000
Separated stems
Total duration 00:00.000
Drums
100%
Bass
100%
Other
100%
Private by design
One model download. No song upload.
When you press the split button, the browser fetches a pinned public HTDemucs model from Hugging Face and verifies it before processing. Your source song, separated stems, mixer settings, and exports stay on this device.
About This Tool
Separate one song into synchronized vocals, drums, bass, and other stems without uploading the audio.
The tool runs a pinned HTDemucs model in a worker, processes overlapping 7.8-second sections at 44.1 kHz, and reconstructs four aligned stems. The model is delivered by Hugging Face only after you press the split button; the song and results remain in browser memory on your device.
Practical tips
- Close memory-heavy tabs before splitting a long track.
- Use WAV when the stems will return to a music editor.
- Compare the original and combined stems before downloading each file.
Common questions
Is my song uploaded to Hugging Face?
No. Hugging Face delivers the pinned public model with a GET request after you start. Your source audio, stems, mixer settings, and exports are never sent there.
Why does the tool need a 166MB download?
The HTDemucs model contains the learned weights used for local four-stem separation. The browser verifies the exact file and caches it when storage permits.
Why are mobile, Firefox, and Safari unsupported?
This release targets the memory, worker, WebGPU, and multithreaded WebAssembly behavior available in current desktop Chrome and Edge. Unsupported profiles do not download the model.