Loudness Guarantee: Your Master Now Always Hits the Target LUFS

One user tapped 👎 with the reason "too quiet" — and they were right: their EDM track came out at −12.5 LUFS instead of the −9 target. We took the case apart, then checked every processing path for the same class of problem. Here's an honest account of what we found and fixed.

The target-LUFS guarantee is now unconditional

Previously, the final loudness normalization was part of the chain configuration. In rare scenarios — a custom module chain, or a stale settings cache in the browser — it could be skipped, and the stages after it (genre EQ, stereo imager, reverb) drifted the loudness away from the goal.

Now the push to the target LUFS happens at the very end of processing, unconditionally — on every mode: regular mastering, custom chains, auto-mastering, batch processing and the Telegram bot. Pick −9, get −9.0. The user's case was reproduced and verified: after the fix their scenario lands exactly on target.

Reference mastering: a major fix

Auditing the adjacent paths surfaced something bigger than the original complaint. Two defects overlapped in the Reference module (matching your track to a reference song):

1. A windowing bug — the spectrum-matching filter was built so that the window function cancelled its core energy. On some track/reference pairs the result lost tens of decibels — down to near silence.
2. Absolute-level matching — if the reference was quieter than your master, or occupied a narrow frequency band, the module pulled the track down to its level, ducking everything else.

Both are fixed: Reference now matches the spectral shape (timbre) of the reference, while your master's loudness is preserved and protected by a true-peak limiter. Regression tests are in place: matching a track to itself keeps loudness intact, a narrow-band reference no longer kills the track, and a bright reference genuinely brightens the spectrum.

Complaints became research cases

When you tap 👎, the original file is now automatically kept for engineering review — together with the genre, reasons and metrics. That's exactly how today's bug was found: on real material, not an abstract description.

Storage is honest and tidy: 30 days, a capped volume with auto-pruning, and the files are used for quality diagnostics only.

Why this matters

Mastering quality is the core of our product. Every 👍/👎 you leave triggers a real process: complaint → saved case → reproduction → fix → regression test so the bug never returns. In the past few days this loop has fired twice — both times starting from a single button tap.

Try it on your track: magicmaster.pro/app — 3 free masterings per day. The user whose complaint led to the fix received 10 tokens — thank you for paying attention!