Improvement to calculation of loudness metadata. #588
Description
This, however, is a low-level bug.
I have models from a highly reputable model author that seem to really not like ultra-low-frequency audio inputs. As a result, they end up getting assigned loudness metadata values that are anywhere up to 4.5dB off. For what it's worth, normalized output for the vast majority of other models seem to give consistent output levels of about -9dBFS. So, mostly working well. (Also some samples of models that are off in the other direction, although I have suspicions about how well they were engineered).
I'm not sure quite what mechanism is at play; but the models in question seem to audibly choke up when presented with what look like DC offsets in the input data, that are actually caused by low-frequencies present in your pink(?) noise input data. When fed a looped-version of the loudness_input.wav, they produce a horrible gurgling sound that seems to correlate to those apparent DC-offsets. And because your input wave file is pink(?) noise, it does have an enormous amount of energy in ultra-low frequency bands (0Hz to 60hz).
It might be a kindness to remove the ultra-low-frequencies from your loudness_input.wav file. Say, frequencies in the range 0hZ to 40- or 50- or 60Hz. Doing so might produce slightly more consistent metadata values. it seems to me that you should be able to do so, although it would require some experimental tweaking to maintain the -18dB(RMS) reference adjustment in ordinary cases.
I'm willing to do the work to generate you a better sample file, and provide guidance on conversion from old to new calibraion procedure, if you would find that helpful. And can (probably) arrange to get you free access to the models in question (they are paid models), if that would be helpful.
Although I don't really have test data that I can run through the NAM workflow to verify that the adjustments produce better results. I CAN, however, re-run the calibration procedure against the collection of models that I currently have, and provide you with data on what changes would occur. Again, if that would be helpful.
Let me know if you need more details, or if I can help in any way.