Stem Lab Hardware Test
iPhone-first · No MMS integration · Your audio stays local
The test did not confirm completion. This can happen when Safari suspends or terminates a tab.
Choose a track
Make a short test clip
No need to trim the song in another app. Start with 5 seconds for the first hardware test; choose a section containing multiple instruments.
Choose a track in Step 1 to activate the clip controls.
Run the lightweight AI test
Controlled diagnostics: run the AI library and 30 MB model load inside a dedicated Web Worker, so an unresponsive inference task does not immediately freeze the page's controls. This does not eliminate memory-related tab crashes. For separation, the page decodes the short clip at 44.1 kHz then transfers raw PCM to the Worker.
Worker not started. Model loading is a separate diagnostic from full separation. No separation is running. Audio decoding happens on the page; AI inference stays in the Worker.
No diagnostic events yet.
Separated audio
This first-stage model separates vocals from instrumental accompaniment; it does NOT split guitars, drums, bass, or synthesizers. On a purely instrumental song, both outputs may be unhelpful. Use a clearly sung passage to evaluate vocal isolation. The WAV comparison below detects duplicate or overly similar signals, not musical quality.
Four-stem Demucs reference
For drums, bass, vocals and other, use the open-source sdemux web demonstration. It runs Demucs on your device, but downloads a substantially larger (~172 MB) model. Upload the WAV clip you just saved in step 02.
Open four-stem reference test ↗Independent third-party site, not part of this ZIP. Phone/Safari compatibility and performance are not verified by us. Stay on that page during inference.
Compatibility report
There is no telemetry, login, or MMS database connection in this package. Your report is only copied or saved when you choose.