I Set a True-Peak Ceiling. Four MP3 Deliveries Still Needed Correction.

I had forty music candidates ready for review. My normalization target was written down: roughly -16 LUFS, with a true-peak ceiling of -1.5 dBTP.

The failed check cost me four additional file corrections before delivery.

Four delivery candidates still needed extra attenuation. Setting the ceiling hadn’t proved that the final MP3 files met it.

The retained reports are from October 10, 2026. I compared the first review encodes with the corrected delivery report, rather than rerunning music generation for this article. The records show exactly which files needed correction and what happened afterward.

I was checking the wrong finish line

These were locally generated music candidates for a small rhythm game. They weren’t approved soundtrack assets yet. I needed reasonably consistent review files so that a louder candidate wouldn’t automatically dominate a listening comparison.

The first technical check covered decoding, duration, loudness, and true peak. Some MP3s decoded successfully while exceeding my chosen peak ceiling. A successful decode answered one question; compliance with my delivery target was another.

Candidate First MP3 true peak Additional gain Final MP3 true peak
02-B -1.33 dBTP -0.47 dB -1.87 dBTP
04-A -0.81 dBTP -0.99 dB -2.00 dBTP
08-B -1.42 dBTP -0.38 dB -2.18 dBTP

These three rows are examples from the four corrected candidates. The values come from the before/after reports. I haven’t reconstructed an unrecorded intermediate waveform to explain every change.

The extra gain and final measured difference don’t match exactly in every row. That’s another reason I read the encoded output instead of treating a gain calculation as the final measurement.

The target wasn’t a result

The FFmpeg loudnorm documentation distinguishes integrated loudness, loudness range, and true-peak targets. It also documents different processing modes. A target is an instruction to processing; my delivered file still needs its own measurement.

My practical mistake was stopping the reasoning at the filter configuration. Encoding came afterward. The review listener received that encoded file, not the values I’d typed into a command.

The before/after reports establish that some initial MP3s exceeded the chosen ceiling and corrected files passed. They don’t isolate every contribution from normalization mode, resampling, and encoding. I don’t have the evidence to blame one codec implementation universally.

I also don’t call -1.5 dBTP a requirement for every game or platform. It was my review-pack ceiling. Another destination can have a different specification.

What the final forty-file report actually says

Final delivery measurement Recorded result
Candidate count 40
Candidates with additional attenuation 4
Integrated loudness range -17.36 to -15.91 LUFS
True-peak range -3.85 to -1.54 dBTP
Highest measured true peak -1.54 dBTP

Every recorded final peak was below my -1.5 dBTP ceiling. That doesn’t mean every track landed at exactly -16 LUFS.

Candidate 04-A illustrates the tradeoff. Its integrated loudness changed from -16.37 to -17.36 LUFS after the additional gain reduction. I accepted a quieter review file instead of claiming both targets were exact.

I’ve published a small measurement extract containing all four corrected rows and the aggregate ranges. It contains candidate IDs and technical values, without private paths or the audio itself.

My delivery checklist starts from the encoded file

I keep the lossless source and the delivery file separate. A failed delivery check shouldn’t force me to regenerate music or repeatedly recompress an already lossy copy.

For this batch, the correction process returned to the lossless source, adjusted the output, and measured the delivery encode again. I preserved both source and output hashes in the private report so I could tell which version had been reviewed.

My checklist is now explicit:

  • Match the expected candidate ID to a real file.
  • Decode the delivery file and check for errors.
  • Measure its duration, integrated loudness, and true peak.
  • Compare those measurements with the chosen delivery criteria.
  • Reprocess only failed candidates, then measure the new output.
  • Keep technical acceptance separate from listening approval.

I don’t count every filename in a directory as a music candidate. The external-drive workflow also encountered ._ metadata files, so the expected candidate list matters. That was an input-selection issue, separate from the audio levels.

Don’t turn a technical pass into musical approval

The October 10 report still marked listening, first-beat, and tempo review as pending. The requested BPM was generation input, not a measured guarantee about the track.

A clean decode doesn’t tell me whether the rhythm is useful. A compliant peak doesn’t tell me whether the opening beat is obvious. Neither tells me whether the candidate belongs in the game.

I don’t have a measured listening-quality gain from this correction, and I won’t invent one. What I have is four failed technical deliveries corrected before treating the pack as ready for that next review.

This belongs beside my embedding threshold experiment: the setting looked reasonable, but the output still needed a test of its own.

I would’ve kept the first encodes if I’d checked only the requested settings. Now the finish line is the actual file I’m about to hand over.

Similar Posts