I have been getting some anomalous results with replaygain. Mostly it works as expected: albums are perceived to be of equal loudness. However, some highly dynamic instrumental albums, and some heavy metal albums are respectively too loud and too soft.

Replaygain auto on Evelyn Glennie - Drumming (Avant-garde classical / drumming / percussion):

Level Adjustment | Peak Level (Adjst)| Filename
----------------------+-----------------------+-------------------------------
+0.00 dB =>+22.97 dB | => -13.2 (+9.74)| Evelyn Glennie - Drumming - 01 - Entrances.mpc
+0.00 dB => +4.39 dB | => -1.71 (+2.68)| Evelyn Glennie - Drumming - 02 - Halasana.mpc
+0.00 dB => +3.54 dB | => -1.22 (+2.32)| Evelyn Glennie - Drumming - 03 - Sorbet No. 1- Latin American Interlude.mpc
+0.00 dB => +6.57 dB | => -0.45 (+6.12)| Evelyn Glennie - Drumming - 04 - Bongo-O.mpc
+0.00 dB => +2.02 dB | => -0.30 (+1.72)| Evelyn Glennie - Drumming - 05 - Sorbet No. 2- Chinese Cymbals.mpc
+0.00 dB => +3.13 dB | => -0.35 (+2.78)| Evelyn Glennie - Drumming - 06 - Prim.mpc
+0.00 dB => +3.78 dB | => -4.92 (-1.14)| Evelyn Glennie - Drumming - 07 - Sorbet No. 3- UDU Trail.mpc
+0.00 dB => +4.45 dB | => -0.28 (+4.17)| Evelyn Glennie - Drumming - 08 - The Anvil Chorus.mpc
+0.00 dB => +6.77 dB | => -3.40 (+3.37)| Evelyn Glennie - Drumming - 09 - Sorbet No. 4- Woodblocks and Falling Instruments.mpc
+0.00 dB =>+10.12 dB | => -0.46 (+9.66)| Evelyn Glennie - Drumming - 10 - To the Earth.mpc
+0.00 dB => +9.35 dB | => -5.06 (+4.29)| Evelyn Glennie - Drumming - 11 - Sorbet No. 5- Wood and metal Chimes.mpc
+0.00 dB => +3.66 dB | => -0.45 (+3.21)| Evelyn Glennie - Drumming - 12 - Pezzo Da Concerto No. 1, Op. 15, for Snare Drum.mpc
+0.00 dB => +1.66 dB | => +0.01 (+1.67)| Evelyn Glennie - Drumming - 13 - Sorbet No. 6- Simtak Debut.mpc
+0.00 dB => +4.28 dB | => -0.45 (+3.83)| Evelyn Glennie - Drumming - 14 - Matre's Dance.mpc
+0.00 dB => +0.00 dB | => -13.5 (-13.5)| Evelyn Glennie - Drumming - 15 - Exits.mpc
+0.00 dB => +3.27 dB | => -0.77 (+2.50)| Evelyn Glennie - Drumming - 16 - Sorbet No. 7- Hi-Hat Playout.mpc
=> +4.37 dB | => +0.01 (+4.38)|

The +4.37 dB album gain makes this drumming album too loud. Does the RMS peak calculation fall short when it encounters this type of highly dynamic music?

Conversely:

Replaygain auto on Creed - Human clay (Post-grunge):

Title | Album |
Level- | (Peak+)| Level- | (Peak+)|
Adjustment| Peak (Adjst)|Adjustment| Peak (Adjst)| Filename
----------+--------------+----------+--------------+---------------------------
-8.18 dB | 40133 (15650)| -8.00 dB | 42725 (17009)| Creed - Human clay - 01 - Are you ready .mpc
-8.07 dB | 40149 (15855)| -8.00 dB | 42725 (17009)| Creed - Human clay - 02 - What if.mpc
-7.87 dB | 41195 (16647)| -8.00 dB | 42725 (17009)| Creed - Human clay - 03 - Beautiful.mpc
-7.14 dB | 42725 (18779)| -8.00 dB | 42725 (17009)| Creed - Human clay - 04 - Say I.mpc
-9.00 dB | 40503 (14371)| -8.00 dB | 42725 (17009)| Creed - Human clay - 05 - Wrong way.mpc
-8.89 dB | 39643 (14245)| -8.00 dB | 42725 (17009)| Creed - Human clay - 06 - Faceless man.mpc
-8.68 dB | 41197 (15166)| -8.00 dB | 42725 (17009)| Creed - Human clay - 07 - Never die.mpc
-6.83 dB | 38461 (17519)| -8.00 dB | 42725 (17009)| Creed - Human clay - 08 - With arms wide open.mpc
-7.71 dB | 38969 (16041)| -8.00 dB | 42725 (17009)| Creed - Human clay - 09 - Higher.mpc
-6.77 dB | 38803 (17798)| -8.00 dB | 42725 (17009)| Creed - Human clay - 10 - Wash away those years.mpc
-6.43 dB | 38953 (18580)| -8.00 dB | 42725 (17009)| Creed - Human clay - 11 - Inside us all.mpc
-7.98 dB | 38057 (15186)| -8.00 dB | 42725 (17009)| Creed - Human clay - 12 - Young grow old.mpc

The –8.00 dB album gain makes this a little too soft, considering that this type of music has to be listened to fairly loud.

I suspect that the RMS energy setting of 95% is what is causing this effect. With loud, continuous energy, highly [studio] compressed pop/rock etc, the replaygain implementation is “turning down” the energy level too much.

With highly staccato, dynamic, percussive solo instrumental, the RMS energy level is not representing the true perceived loudness of the music, and replaygain tends to “push” it too high.

If you were to analyse the histograms of the RMS energy levels for “Drumming” you would probably find the most common RMS value being quite soft, and even the 95% RMS value would be quite low.

What I’m suggesting is that the RMS value needs to be set on an album by album basis. in the range of ~70% up to ~99%. Perhaps the spread function of the histogram could be used to set this value. (The histograms for three types of sound can be seen here: http://privatewww.essex.ac.uk/~djmrob/repl...al_process.html.)

I’m not sure how you would do this, but some analysis of the RMS histogram might be used to calculate a value for RMS peak.

Also one thing is that if clipping protection is in use, then the gain adjustment may not have the desired effect, because the clipping protections kicks in.

But, if you don't use clipping protection, there can be audible clipping especially after some gain adjustment.
I suggested that Andree implements a hard limiter like in Vorbis plugin, but I don't know if it will happen. Of course hard limiter is probably harming quality some, but in some situations it could be useful. With hard limiter you can avoid clipping and too quiet sound.
If Andree is not going to implement hard limiter, maybe someone else will (*cough*NickSD?*cough* )

Frank has said that he will add album gain adjustment soon.

Ruse,

I'm collecting Replay Gain problem samples like this.

If possible, can you email me 1 track from the drumming album please? Preferably no larger than 5MB-10MB - the quality doesn't really matter. (If you're not on the kind of connection where you want to transfer large files, then don't worry about it.)

Cheers,
David.
http://www.replaygain.org/

2Bdecided wrote:


Sure.

JohnV wrote:

If this were the case, JohnV, then the drumming album would sound too soft. I am concerned that it sounds too loud. If clipping protection was engaged, it would be smoothing off the drum stick attack transients. That is not what is happening. The dynamics of the music are OK, but the whole album has been pushed in volume too much by replaygain, compared to other albums with more average dynamics.

To clarify just a bit, I'm suggesting that "Drumming" might have come out at a more representative average volume if the RMS peak had been set at about 98-99%, and "Human Clay" might have come out at a better album volume if RMS peak had been set at 70-80%.

Hi Ruse,

Thanks for sending the file to me. What you've said above pretty much sums up the problem. However, it's not the percentile that needs changing, it's the RMS averaging time. I found a similar problem when first developing ReplayGain, with a track containing four castanets clicks. Whereas ReplayGain calculates the RMS energy over 50ms, all the energy of the drum hits is concentrated in approx. 1-5ms bursts - averaged over 50ms this appears quieter than it really is, so Replay Gain makes the track louder to compensate.

It's not a great idea to alter the parameters of the ReplayGain calculation on an album by album basis - if you have to do this, then you might as well just change the gain by hand!

However, it is possible (probable even) that the 50ms RMS window isn't an ideal length, and should be changed. If it's made too short, ReplayGain fails for many many audio clips - however, I didn't extensively test the threshold of this, and it could be that a shorter value would solve the problem with your drumming clip, without causing problems elsewhere. It's probable there's no ideal solution (ReplayGain is a simple solution to a complex problem, so it's not going to be perfect), but I think things can be improved.

Frank Klemm has made the following change, which may help with your file:



I haven't tried this modification yet, but it's great that people are improving Replay Gain like this - I know Frank has put in a lot of work, and still has some great ideas to implement.

With care, any modifications to Replay Gain are compatible with the first version: files will still be brought to the same "average" loudness, and it's only files which were causing problems before that will yield different (better) results with the newer versions. (I can't guarantee this, but that's how Replay Gain is designed, and you'd have to intentionally mess it up to cause incompatibility problems!)

Thanks for sending me this clip. Watch Franks web site for a release of the modified code.

Cheers,
David.
http://www.replaygain.org/

This sounded very hopeful 2Bdecided, until I tried Frank's new replaygain 0.81c on test track as follows:

Level Adjustment | Peak Level (Adjst)| Filename
----------------------+-----------------------+-------------------------------
+6.57 dB => +7.36 dB | -0.45 => -0.45 (+6.91)| Evelyn Glennie - Drumming - 04 - Bongo-O.mpc
=> +7.36 dB | => -0.45 (+6.91)|


The track was previously processed with replaygain 0.81a and suggested a level adjustment of +6.57 dB, which was too loud. The new replaygain has suggested a level adjustment of +7.36 dB, which is even louder!

Frank says on his site:

but I'm not sure his adjustments have solved this drumming transients problem: in fact it appears to have exacerbated the problem.

He's still tweaking...

I think it'll improve - he's aware that modification version 1 didn't quite hit the spot!

Have you sent him that track? If not, I will.

Cheers,
David.

Sending....

New version replaygain 0.81d test:

Replaygain 0.81d © 2001-2002 Klemm/Robinson/Sawyer

Level Adjustment | Peak Level (Adjst)| Filename
----------------------+-----------------------+-------------------------------
-8.18 dB => -7.95 dB | +1.76 => +1.76 (-6.19)| Creed - Human clay - 01 - Are you ready .mpc
-8.07 dB => -7.89 dB | +1.76 => +1.76 (-6.13)| Creed - Human clay - 02 - What if.mpc
-7.87 dB => -7.59 dB | +1.99 => +1.99 (-5.60)| Creed - Human clay - 03 - Beautiful.mpc
-7.14 dB => -6.97 dB | +2.30 => +2.30 (-4.67)| Creed - Human clay - 04 - Say I.mpc
-9.00 dB => -8.90 dB | +1.84 => +1.84 (-7.06)| Creed - Human clay - 05 - Wrong way.mpc
-8.89 dB => -8.71 dB | +1.65 => +1.65 (-7.06)| Creed - Human clay - 06 - Faceless man.mpc
-8.68 dB => -8.34 dB | +1.99 => +1.99 (-6.35)| Creed - Human clay - 07 - Never die.mpc
-6.83 dB => -6.75 dB | +1.39 => +1.39 (-5.36)| Creed - Human clay - 08 - With arms wide open.mpc
-7.71 dB => -7.47 dB | +1.51 => +1.51 (-5.96)| Creed - Human clay - 09 - Higher.mpc
-6.77 dB => -6.62 dB | +1.47 => +1.47 (-5.15)| Creed - Human clay - 10 - Wash away those years.mpc
-6.43 dB => -6.16 dB | +1.50 => +1.50 (-4.66)| Creed - Human clay - 11 - Inside us all.mpc
-7.98 dB => -7.81 dB | +1.30 => +1.30 (-6.51)| Creed - Human clay - 12 - Young grow old.mpc
=> -7.80 dB | => +2.30 (-5.50)|


The left most column represents gain applied by replaygain 0.81a, and the next column 0.81d. There is slightly less negative gain applied by latest version, albeit inaudible (3 dB is minimum threshold of gain detection by ear). However, the latest replaygain is applying less negative gain to this music, which in my opinion, is heading in the correct direction. That’s great.

“Drumming”

Replaygain 0.81d © 2001-2002 Klemm/Robinson/Sawyer

Level Adjustment | Peak Level (Adjst)| Filename
----------------------+-----------------------+-------------------------------
+22.97 dB =>+22.40 dB | -13.2 => -13.2 (+9.17)| Evelyn Glennie - Drumming - 01 - Entrances.mpc
+4.39 dB => +5.50 dB | -1.71 => -1.71 (+3.79)| Evelyn Glennie - Drumming - 02 - Halasana.mpc
+3.54 dB => +4.80 dB | -1.22 => -1.22 (+3.58)| Evelyn Glennie - Drumming - 03 - Sorbet No. 1- Latin American Interlude.mpc
+6.57 dB => +7.82 dB | -0.45 => -0.45 (+7.37)| Evelyn Glennie - Drumming - 04 - Bongo-O.mpc
+2.02 dB => +2.41 dB | -0.30 => -0.30 (+2.11)| Evelyn Glennie - Drumming - 05 - Sorbet No. 2- Chinese Cymbals.mpc
+3.13 dB => +3.46 dB | -0.35 => -0.35 (+3.11)| Evelyn Glennie - Drumming - 06 - Prim.mpc
+3.78 dB => +4.70 dB | -4.92 => -4.92 (-0.22)| Evelyn Glennie - Drumming - 07 - Sorbet No. 3- UDU Trail.mpc
+4.45 dB => +5.11 dB | -0.28 => -0.28 (+4.83)| Evelyn Glennie - Drumming - 08 - The Anvil Chorus.mpc
+6.77 dB => +5.43 dB | -3.40 => -3.40 (+2.03)| Evelyn Glennie - Drumming - 09 - Sorbet No. 4- Woodblocks and Falling Instrum
ents.mpc
+10.12 dB =>+10.58 dB | -0.46 => -0.46 (+10.1)| Evelyn Glennie - Drumming - 10 - To the Earth.mpc
+9.35 dB => +9.28 dB | -5.06 => -5.06 (+4.22)| Evelyn Glennie - Drumming - 11 - Sorbet No. 5- Wood and metal Chimes.mpc
+3.66 dB => +4.35 dB | -0.45 => -0.45 (+3.90)| Evelyn Glennie - Drumming - 12 - Pezzo Da Concerto No. 1, Op. 15, for Snare D
rum.mpc
+1.66 dB => +2.14 dB | +0.01 => +0.01 (+2.15)| Evelyn Glennie - Drumming - 13 - Sorbet No. 6- Simtak Debut.mpc
+4.28 dB => +5.30 dB | -0.45 => -0.45 (+4.85)| Evelyn Glennie - Drumming - 14 - Matre's Dance.mpc
+0.00 dB => +0.00 dB | -13.5 => -13.5 (-13.5)| Evelyn Glennie - Drumming - 15 - Exits.mpc
+3.27 dB => +3.66 dB | -0.77 => -0.77 (+2.89)| Evelyn Glennie - Drumming - 16 - Sorbet No. 7- Hi-Hat Playout.mpc
=> +4.88 dB | => +0.01 (+4.89)|

The title gain applied to Evelyn Glennie - Drumming - 04 - Bongo-O.mpc with each version:

0.81a - +6.57 dB
0.81c - +7.36 dB
0.81d - +7.82 dB

In my opinion, this is heading in the wrong direction for this type of music.

It would appear that whatever Frank’s doing to the algorithm, it’s applying more gain (less –ve gain) across the board:

Album title------------ Album gain
--------------------0.81a-------0.81d

Human Clay_____-8.00 dB___-7.88 dB
Drumming______+4.37 dB___+4.88 dB

Also, recursive isn't too good right now... it treats all files like they're part of the same album. Should have an option to do that based on the folders the mpc's are in.

See this thread for some batch files that will ensure files are seperated into albums by their folders:
http://www.hydrogenaudio.org/forums/showth...=&threadid=1041

in b3.bat you could also add "--smart" to the command line and then replaygain would only give a replaygain value to files that don't already have it, but I noticed in the replaygain source code there was a note that said --smart macht Unsinn (makes nonsense) so there might be some bug with it, i'm not sure.

Thanks Leto...

BTW: I'd like to know what happened to your last Duncan ;p