Published by PitchChanger · Updated · How these guides are made
How to use this workflow
Build a controlled comparison
Import the untouched source and individual estimated roles at a common origin with tempo adaptation disabled. Use an external editor to sum the roles once. Do not include backing mixes that already contain some of the same roles. Compare sample rates and check whether downloads have independent output gains.
A one-frame counterexample
For a synthetic sequence x = [0, 0.5, 0, −0.5, 0], a one-frame delayed copy is y = [0, 0, 0.5, 0, −0.5]. The difference x − y is [0, 0.5, −0.5, −0.5, 0.5]. The content is merely delayed, yet the subtraction is large. This arithmetic example does not measure any real stem model.
Swipe or scroll sideways to read all columns.
| Frame | Source x | Delayed y | x minus y |
|---|---|---|---|
| 0 | 0 | 0 | 0 |
| 1 | 0.5 | 0 | 0.5 |
| 2 | 0 | 0.5 | −0.5 |
| 3 | −0.5 | 0 | −0.5 |
| 4 | 0 | −0.5 | 0.5 |
Interpret what remains
After alignment and gain matching, residual energy shows a waveform difference under those conditions. It does not identify which instrument estimate is useful for practice. A sum can reconstruct well while individual roles contain leakage; conversely, independently scaled downloads may fail a strict null without indicating a broken separation.
Write a reproducible result
State files, offsets, gains, rate and summing method. Keep both the residual and a normal listening comparison if you perform the test. Do not describe the example above as an actual null test. PitchChanger offers named WAV exports and preview, not a source-minus-sum analysis button.
Common questions
Does a perfect sum prove clean individual roles?
No. Leakage can cancel or redistribute when roles are combined. Check the individual role against the musical task as well.