For the next few days I’m going to be running some audio tests – repeating songs and blocks from time to time. The songs/blocks have different settings/adjustments and I’ll be recording the streams on my end and then analyzing them using AI.

There are 3 areas I’m working on – Replay Gain which is basically loudness normalization. With the variety of music played, a rough estimate of the “loudness” difference between songs (soft to loud) in the streaming library is about 10-15 dB. The basic idea is to set up a baseline where you can find a middle and run from that. I’m going to be setting up 3 different baselines which is an experiment but I think it will work.
(ETA – Didn’t need to, I ended up working with -14 across the whole library)

Headroom – The current Replay Gain settings you’re hearing right now are in the middle of the three levels I’m looking at (I want to go a bit higher for most of the library), and using them I was able to cut the true peak headroom down to .5dB which is pretty tight.

Peaking/Clipping- At the current settings the headroom is pretty good for most of the songs and increasing it would be a mistake, however about 2.5% of the library is too hot (rough estimate) and is going over the headroom. Lowering the Replay Gain level on those will help out significantly but there will still be some songs that need manual corrections and I’m working on that.

I’m finding it pretty interesting, I’ve been using AI to scan the library – this is what it looks for:

1. Integrated loudness (LUFS)
2. True peak level (dBTP)
3. Peak headroom
4. Amount of gain needed to reach the target LUFS
5. Whether that gain would push the track above the true-peak ceiling
6. Amount of peak reduction needed to prevent clipping
7. Tracks that are unusually loud
8. Tracks that are unusually quiet
9. Overall loudness range across the library
10. Distribution of tracks above and below the target
11. Files that cannot be analyzed correctly
12. Missing, damaged, or unreadable files
13. Before-and-after values for tracks selected for correction
14. Confirmation that corrected tracks meet the loudness and peak limits

ChatGPT-5.6 Sol.

And then when I upload the recorded blocks to AI it checks:

1. Integrated and short-term loudness (LUFS)
2. True peak levels (dBTP)
3. Clipping or near-clipping
4. Dynamic range
5. Crest factor
6. Song-to-song volume differences
7. Changes to loud transients
8. Limiter activity and peak reduction
9. Effects on quieter passages
10. Differences introduced by the broadcast chain and MP3 encoding

ChatGPT-5.6 Sol.

It’s all numbers – AI isn’t looking at wave forms, analyzing the actual music or anything like that. It’s just looking at numbers and measurements.
– David