Skip to content

Stem Quality Judgment

Flag: --choose-stems

Suno exports numbered stems, but their quality varies — some have audible bleed, phase issues, or missing frequency content. When --separate-missing generates AI stems via Demucs or UVR, you now have two versions of each stem. This feature helps you pick the better one.

What it evaluates

  • Is Suno's provided stem good enough? Compares RMS energy, spectral centroid, and silence ratio to detect thin or empty stems.
  • Is the Demucs output better than the original? Computes normalized cross-correlation between both versions, then scores on silence ratio and energy to recommend the cleaner one.
  • Replace one stem or all? Reports per-stem with individual confidence scores, so you can swap only the bad ones instead of throwing out Suno's work wholesale.

When to use it

  • After running --separate-missing to generate AI stems alongside the originals
  • When you hear bleed or phasing in a specific stem and want a data-backed second opinion
  • When you want to automate "listen to both, pick the better one" across a batch of songs

Example

1
2
3
4
5
# Generate AI stems, then compare
suno-to-ableton process /path/to/my-song --separate-missing --choose-stems --apply

# Or standalone after processing
suno-to-ableton choose-stems /path/to/my-song --apply

What --apply does

Without --apply, prints a comparison report. With --apply, copies the recommended stem into the output directory, replacing the weaker version.