R128Norm, A simple normalizer based on EBU R128 Options

[kode54]

Welp. I went and made a normalizer DSP out of libebur128. Have fun with it. Oh, and I posted the source, which requires the libebur128 files from the R128Scan source.

It's currently hard coded to maintain a preview latency of 500ms, and uses the short-term loudness level measured from a three second window as the gain correction reference, which it updates every 100ms. I may make this configurable in the future.

Change log:

• Version 1.10 - 2011-02-07 14:12 UTC:
  
  • Fixed a bug in libebur128 when handling weird sample rates.
• Version 1.9 - 2011-02-05 06:11 UTC:
  
  • Re-enabled increasing the gain level.
  • Changed momentary and short-term loudness to gated loudness.
• Version 1.8 - 2011-02-04 13:44 UTC:
  
  • Disabled the gain from increasing above 1.0, for now.
• Version 1.7 - 2011-01-28 06:19 UTC:
  
  • Fixed initial gain values on startup, again.
• Version 1.6 - 2011-01-27 23:43 UTC:
  
  • Now correctly flushes the buffer on playback termination.
• Version 1.5 - 2011-01-27 23:07 UTC:
  
  • Reenabled momentary loudness polling.
  • Changed volume ramping to 1 dB every 50ms.
  • Current scale is now forced to the detected target scale after the initial buffering completes.
• Version 1.4 - 2011-01-27 20:19 UTC:
  
  • Increased the latency to maintain at least 500ms worth of samples.
• Version 1.3 - 2011-01-27 20:06 UTC:
  
  • Reverted short-term gain level changes to instantaneous again.
• Version 1.2 - 2011-01-27 19:58 UTC:
  
  • Disabled momentary loudness polling and increased short-term loudness polling frequency.
• Version 1.1 - 2011-01-27 19:22 UTC:
  
  • Adjusted gain level changes a bit.
• Version 1.0 - 2011-01-27 19:00 UTC:
  
  • Initial release.
• Version 0.0 - 2011-01-27 16:53 UTC:
  
  • Created project.

[markanini]

Seems useful for night listening among other things. Thanks!

[mudlord]

Oh yeh, its around 2.8x faster than my port of Garf's Dynamics Compressor.

So, its quite nice, really.

[GeSomeone]

its around 2.8x faster than my port of Garf's Dynamics Compressor.

They don't do the exact same thing, as you will know. The effect of foo_r128norm is more like that of Vlevel, i.e. it seems to adjust the volume (slowly) over time as opposed to compressing the dynamics to make it sound louder. It sounds obviously radio-ish. The calculations will probably be much different (from Vlevel) though, and a fixed 3 sec. sliding window is used.

It could have it's use, configurable settings are welcome. It might need something like a maximum gain or lowest (source) level to amplify (threshold). Right now start of (next) track after a couple of seconds silence and some fade-outs are handled unpleasantly.
A name suggestion? foo_dsp_r128norm .

This post has been edited by GeSomeone: Feb 1 2011, 23:46

[kode54]

When I add configuration, I will add an option to only reduce gain instead of also increasing it.

[TuNk77]

Thank you for making this DSP kode54, now I can finally get rid of that horrible Vlevel.

[markanini]

I don't mind that it raises volume of quiet parts personally but it gets extremely loud right at the start of tracks. This becomes especially annoying when listening to live recordings. I guess I'm reporting a simmilar issue as GeSomeone. Perhaps adding buffer och that's capable of looking ahead into the next track would be a solution for the issues I'm experiencing. FWIW

[kode54]

It already looks ahead, but it doesn't ask for track change notifications, so it keeps whatever amplification level was achieved by the end of the previous track.

[markanini]

That would explain the loud blasts during track transitions.

[kode54]

It should not be blasting, because it adjusts the gain level fairly quickly, and it looks ahead by half a second.

[amalone]

It should not be blasting, because it adjusts the gain level fairly quickly, and it looks ahead by half a second.

I noticed momentary blasts also. I think it happened with Last FM with songs from different artists. I believe it also happened when playing songs on a playlist with Replaygain shut off.

A great component.

[markanini]

Here's an excerpt where output clips at a track transition: http://www.box.net/shared/0qpuv6r21f
From 0:18 I start the second track directly and gain is more reasonable.

[kode54]

Hopefully, locking the gain scale to 1.0 or less will fix that problem. I can't make the component request track change marks because it could cause gaplessness issues when using resampling DSPs and such.

[GeSomeone]

Hopefully, locking the gain scale to 1.0 or less will fix that problem.

Version 1.8 highlights: Disabled the gain from increasing above 1.0, for now

Yes, it does ... but now it is just a limiter instead of a normalizer. Wouldn't the correct solution be to make it work "gated", meaning ignore the silence (and probably also near silence like tape noise etc)? Somewhat like you did in the R128scanner.
Another possible approach could be looking ahead just as much as looking back. e.g. look back 1.5s and look ahead 1.5, still a 3s sliding window :-) Just adding suggestions to the pile.

[amalone]

That change certainly made School from Crime of the Century by Supertramp much better. With version 1.7, it would get almost deafening in parts as it is a very dynamic song. It did not seem to change anything with the flac file previously posted though. Might that be because the quiet part is so long?

Thanks for working on this.

[markanini]

I agree with GeSomeone. With the lastetst update R128Norm loses functionality that makes it a gem of a volume leveler for me. Please consider making it selectable. And if convenient please add track change awareness, also selectable.

This post has been edited by markanini: Feb 5 2011, 00:40

[kode54]

Try now. Although I had to make some ugly additions to the library so I could use EBUR128_MODE_I indefinitely without allocating more and more memory.

[amalone]

I installed 1.9 and tried it. At around the 12 second mark the song School i mentioned before blasts pretty strongly. That part of the song was recorded to go from near silence to quite loud instantly. I did not try the blast flac that was posted before. Thanks.

[markanini]

It's better, blasts are reduced even though they're not avoided completely. Volume leveling sounds more natural than before. Thanks for your great work so far!

[GeSomeone]

Thanks very much for your work, I'm sorry it made the code ugly.
What you changed, did help, but unfortunately, not a lot. If a track ends really abrupt the next track may start at a reasonable volume. However when a track fades out a bit slower, R128norm is cranked up at the start of the next track (the "blast" effect). Is the look ahead you mentioned is working as intended?

This happens not only at track changes. I uploaded a sample of the intro of Monkey Wrench by Foo Fighters from their Greatest Hits. There is a 1.15s break at 0:11.5 that triggers the "blast" effect. At least when played from the beginning, the strange thing is when I start it and skip to 0:10 immediately it won't happen.

In this same sample after seeking (with the seekbar) to 0:09 the volume becomes instantly very high as well. When doing so to 0:08 or 0:10 this doesn't happen.

Another example I will just mention is the intro of Echoes by Pink Floyd. R128norm (used as only DSP) introduces severe clipping there. That can't be right.


used foobar 1.1.2 with track gain applied and only active DSP "EBU R128 compressor" (1.9)

[kode54]

Updated.

This DSP is not guaranteed to prevent clipping. Use the Advanced Limiter if you also need that.

[amalone]

Installed 1.10 and I think it blasted louder than the previous versions. So far, version 1.8 has been the only one that has not done it. ReplayGain on and Advanced Limiter on. Thanks.

This post has been edited by amalone: Feb 7 2011, 20:13

[kode54]

1.8 didn't increase the gain at all, so it was only a limiter.

[amalone]

I understand. Well, the song "School" really blasted with version 1.10. It was louder than with 1.9. I think I heard it on streaming radio also. Thanks.

[GeSomeone]

This DSP is not guaranteed to prevent clipping.

Ah, it's just that every other dynamics compressor I've come accross (by far not all that exist) did have a mechanism to keep the signal (just) below 0dBFS. Maybe the term is commonly used for a compressor/limiter combination.

I hope you have an opportunity to play with that sample (and others) to find out what's going on.