Spaces:

jeonchangbin49

/

De-limiter

like 19

Running

Hi, endelways! Thank you for your interest. What a coooool song!

First of all, I would like to notify that my model is not a de-clipper, this is a De-limiter, the inverse processor of a limiter. The goal of a de-limiter is not just restoring clipped regions, but restoring the overall dynamic of a song. So it may increase the overall volume and loudness gap between the quieter and louder parts of music.

Let me simplify your song as 4 parts, "Verse1 - Chorus1 - Verse2 - Chorus2". I have checked the waveform of your song and the Verse1 and Verse2 parts doesn't seem heavily compressed unlike 2 chorus parts. It may result in the quite large loudness difference of the de-limiter output of your song. Just curious... did you use different limiters (maybe different threshold settings?) on Verse and chorus?

I think the verse and chorus use the same limiter with the same settings, but then, after limiting, the volume of most of the instruments in the verse is reduced in the mixer. So the verse is quieter than the chorus, but its peak volume is the same as in the chorus because there are instruments whose volume has not been reduced.

I was able to increase the strength of de-limiting when I cut Verse 2 out of the track and processed it with your method separately.

Cut Verse 2:

The processed Verse 2 if it is processed as part of the whole track:

Processed Verse 2 if it is processed separately:

Also, to give a clearer indication of the improvement, I've inserted the separately processed Verse 2 into the full processed track.

Processed full track:

Processed full track with inserted Verse 2 processed separately:

Files: https://drive.google.com/drive/folders/1O6E4A-qOq8r2hSoHFLopz5uJTXkPGbAN?usp=sharing

I'm assuming that your de-limiter is adjusted for the whole track, based on its loudest parts, but then it may not be appropriate for its quieter parts. Is it possible to modify the method so that it detects the quieter parts and adjusts to them accordingly?

You are right. Louder parts affect quieter parts, and vice versa. I also personally like the last one that you gave (separately processed verse 2) but I don't think this will always be better than processing a whole song at once. As I said, the goal of a de-limiter model is increasing the overall dynamic range. It is natural that the gap between quiet part becomes more quieter than louder parts because that is a state before a limiter applied.

I think in your case, processing each part separately is the great option, although it needs slightly more effort. Maybe the "Parallel Mix" can also be the great option. Have you tried it already?

I like the raw processed audio better, it's more vibrant 😊

I have also encountered another problem. The method doesn't work on one of the songs in my collection. Can you please check why?

https://drive.google.com/drive/folders/12eAuw1mnB6qt7gU1ce9Osu3KzApP5BTO?usp=sharing

For all my other tracks your method works well, great job!

Hmm... that's weird. I've never seen this type of problem, and honestly, I don't know why. Maybe similar type of song is not contained in our training dataset? Maybe because kicks are missing in the original song's chorus parts? Not sure. I'll let you know when I get some idea on this.

Thank you for looking into it. Good luck with your future research!

Found a similar problematic example: https://drive.google.com/drive/folders/1MGix75Il5qJILi1EjHQGwQjrvyxD8zCb?usp=sharing