Published by PitchChanger · Updated · How these guides are made
Hear the difference
These are known authored sources and their sum, not model outputs.
Pause your workspace preview before comparing these clips.
Original synthetic material. The settings below distinguish authored inputs from processed outputs; no performer or physical-device test is implied.
How to use this workflow
Choose a meaningful short excerpt
Open Stem Splitter with a clip under six minutes that includes both exposed and overlapping instruments. A compatible WebGPU browser is required.
Inspect one category at a time
Run Split four stems. Use solo and mute to examine Vocals, Drums, Bass and Music · Other instruments. Compare with your unchanged source in a separate player.
Judge the intended use
Check the exact entrance, sustained notes and transitions needed for practice. Export the named WAV or mix you need, then verify the saved file before discarding any working copy.
Reproduce the ambiguity with three sample positions
For two unknown sources A and B, the observed mix is A + B. Add a small value d to A and subtract the same value from B: the sum stays unchanged. The numbers below are an original arithmetic demonstration, not outputs from Demucs and not a quality score.
Swipe or scroll sideways to read all columns.
| Known A | Known B | Same mix | Different possible sources |
|---|---|---|---|
| 0.20 | 0.10 | 0.30 | 0.25 + 0.05 |
| 0.30 | −0.05 | 0.25 | 0.40 + (−0.15) |
| 0.00 | 0.20 | 0.20 | −0.10 + 0.30 |
Compare a mixture with the sources that actually made it
The downloads provide a 220 Hz source, a 440 Hz source and their sum. Each is two seconds at 48 kHz. They are authored signals before mixing, not AI-isolated examples. Knowing their construction lets you compare true source material with the mixed signal without inventing a separation result.
A real model uses learned musical patterns to choose plausible outputs from a much harder problem. Vocals and instruments overlap in frequency and time; reverb and processing can spread one source into other categories. No fixed frequency boundary assigns every sound to an instrument.
Use three checks on an estimated stem
First inspect bleed: can you hear an unwanted source during a gap in the desired part? Next inspect loss: does the desired note, consonant or drum attack disappear? Finally inspect artifacts: do sustained sounds flutter, smear or change texture compared with the original?
A stem useful for locating bass notes may still be unsuitable as an exposed production track. Evaluate it in the context you intend to use. Raising the monitoring volume can make spill easier to notice; it does not restore missing information.
What this browser workflow can supply
The current model exports four category WAVs and named mixes such as Drumless, Bassless, Instrumental and Acapella. Music · Other instruments is a broad remaining category, not a dedicated guitar or piano track. There is no six-stem selector, CPU fallback or per-stem gain slider.
For adjustable balances, export the individual WAVs to an external DAW. If you need the actual original tracks, request the multitrack session from its owner. Do not describe an estimated Acapella as an unprocessed studio vocal.
Common questions
Does summing the stems prove they are the originals?
No. Even an exact sum can come from many different assignments of sound between stems. Compare the individual parts against known originals when those are available.
Will every separation have audible damage?
The amount depends on the source and task. Artifact-free recovery is not guaranteed; inspect the saved part where you actually need to use it.