How to use Audio Stem Separator
- Choose a mono or stereo PCM16 WAV, no longer than 15 seconds. Complete browser verification before the recording is uploaded to the server worker.
- Select Separate and preview stems. Listen to vocals, drums, bass and other; check whether wanted sounds are missing or bleed into another stem.
- Use Export current result to download the four WAV files before the job expires after 15 minutes. Keep the source recording for comparison.
Example: Audio Stem Separator
Separate a short voice, bass and percussion recording into stems you can review.
Supported inputs and limits
- Server-side Demucs 4.0.1 HTDemucs checkpoint.
- One PCM16 WAV, mono/stereo, 8–48 kHz, up to 20 MiB and fifteen seconds.
- Mono is duplicated before inference; output is 44.1 kHz stereo PCM16.
- Four estimated sources, not left/right splitting.
- No GPU assumed.
More details
Source bleed and artifacts are expected; shared gain prevents clipping.
One worker task at a time.
Jobs expire after fifteen minutes.
Where your input is processed
Your file is uploaded to an isolated server worker after Cloudflare human verification. The verification request does not include the file. Inputs are removed after processing. Downloads belong to this browser session and expire after 15 minutes. The worker removes expired jobs automatically; if it stops, temporary files can remain until cleanup resumes or the scheduled fallback runs.
Listen to each stem and then the combination
A sound can appear in more than one estimated stem, and a quiet voice or transient may be lost. Check the part you want in isolation, then listen with the other stems before using it in an edit. If any estimated sample would clip, the worker applies the same gain reduction to all four outputs. Their combined level can therefore differ from the source. Keep the original recording and make sure you have permission to reuse it.
Questions about Audio Stem Separator
Is this stereo-channel splitting?
No. The trained model estimates vocals, drums, bass and other from the mixture.
Will sources be perfectly isolated?
No. Instruments and voice can bleed into other stems or be damaged.
Why is the clip short?
CPU inference is resource-intensive, so this initial worker bounds duration and concurrency.