I know this is essentially a repeat thread, but I feel that the last thread containing this discussion was originally a somewhat different discussion altogether. I feel that this is a pretty important issue to sort out, so here is my revised proposal for everyone to consider:

Instead of replaygaining MP3s down to 89dB with MP3Gain (after encoding), why not run a wavegain analysis on the raw WAV files, then plug in the --scale value into LAME?

A Number of benefits:

• This gives 89.0dB precisely, not +- 1.5dB like MP3Gain does.
• No lossy dithering is necessary, as would be if we simply wavegain the raw waves before encoding.
• For modern, VERY LOUD recordings, LAME tends to allocate bits more freely. This seems to limit that behaviour. Many modern recordings are thus ~10% smaller.

The debate is really over the third point. What is getting thrown out exactly? As 2BDecided suggested in the previous thread, LAME's ATH model may not like this, since it might be assuming peaks are at 103dB (which they won't be near on compressed and normalized tracks).

Ultimately it comes down to whether the difference is audible. I can't personally ABX a difference, but my ears aren't that well tuned. I have encoded ~30 albums so far with both methods (using --alt-preset standard), in several genres, so if anyone wants samples uploaded please let me know.

1.5dB is meant to be unnoticeable yet a safe measure to apply gain over on MP3's.

It has been said that the average person can only perceive volume changes of 3dB.

So my advice is to just relax, MP3Gain is good enough.

That was the least of my concerns Neo. I am more interested in the fact that overcompressed, OVERLOUD recordings are bitrate hogs, and this seems to be a solution.

I'm not putting that long reply into this thread too!

Dibrom et al - your advice is needed in one thread or another!


Cheers,
David.

@Jebus

You don't seem to be getting much response re ABX tests. If you post some samples somewhere, clearly showing which is which so I can grab the right pairs to download, I can have a crack at some tests if you wish. It would be be best if the samples are ready to go as mp3s, rather than a wav that I need to encode/scale/whatever. If you present them in their final form, I'll check them out.

Best to keep the samples relatively small (15 - 30 seconds a piece) if possible, rather than entire songs. I'd also be interested in some samples where there is a lot going on, and some with a solo instrument, or the beginning of a song. ( I have my reasons for wanting both, but will explain more if my findings match my expectations after I ABX them.)

Den.

If you're comparing wav>lame>mp3gain vs wav>wavgain>lame, then the mp3gain and wavgain versions could be different by up to 0.75dB. You might ABX this on loudness alone.

If you are planning to ABX, do a "fake" comparison where you use exactly the same attenuation as suggested within mp3gain, so there's no level difference between the files.

Cheers,
David.

P.S. Lame devs - still looking for your advice in the other thread:
http://www.hydrogenaudio.org/forums/index....1&t=10637&st=39

Okay, here are some samples to mess with.

All are done with LAME 3.90.3 --alt-preset standard, and with the exact same binary at that. As 2Bdecided suggested, their gains match PERFECTLY, so there is no volume difference at all the mess up the ABX tests.

Here is a Floyd track, since everyone here likes floyd. It's pulled off the Echoes compilation because most of their other stuff isn't compressed that much, so there is no point.

One of these Days (MP3Gained) bitrate ~212
One of these Days (--scale 0.5909) bitrate ~206

Here is a Ministry track off Animositisomina, one of the LOUDEST discs I have ever encountered (as you can see from the --scale value used). Look at the bitrate savings!

Impossible (MP3Gained) bitrate ~270
Impossible (--scale 0.2515) bitrate ~230

Like said in the General forum, I had a similar issue here
Please repeat that encoding with the -Y switch, if the bitrate differnece goes down noticeably, it's the SFB21 (very high freq)/mp3 VBR problem.

Yup, you appear to be right. I put results in the other thread (shouldn't have started this one... mods want to lock it?)