Loudness that survives re-encoding

2026-10-22 · 4 min read

Most loudness bugs are introduced by measuring too early. If you normalise before the final lossy encode, the codec's own gain behaviour moves you off target — usually by 0.3–0.8 LU on 24-bit sources.

What we changed

The pipeline now measures integrated loudness on the decoded output of the last encode pass, and only re-runs the chain if the deviation exceeds 0.2 LU. Re-runs are capped at two.

Cost

Roughly 6% more CPU per job; first-pass success went from 94.1% to 98.7%.

See the job API or the changelog for the release notes.