MPC VBR flaws (low volume & ringing), audible under specific conditions Options

[guruboolez]

I’ve read recently some complaints about musepack and distortions occurring with classical music (examples here and here). There were no ABX tests to confirm them. According to my previous listening tests at ~175 kbps, musepack performs not only very well with various kinds of instrumental and vocal samples, but also better than competitors. But I’ve also noticed in the past one issue with this audio format that my previous test didn’t revealed, and it’s a very big one. I’d like to bring out this problem to the community, which wasn’t as far as I know warned about this kind of flaw.

Before carrying and before some seeing zealous users bare its teeth, I have to make clear that this issue only occurs in specific conditions. The problem is confined to low-volume musical content, and is mainly audible when this content has to be listened to a higher playback volume. In other words, affected tracks must have low volume parts, and tracks with high dynamic are not really concerned (you can’t constantly push the volume on such material: your neighbors won’t appreciate it). The problem becomes really critical with low-volume tracks only. People who have to live with the consequences of the “loudness war” are certainly not used to encounter such tracks, but for classical fans, tracks that are replaygained at +10 dB, +20 dB and sometimes +30 dB are all except a rare thing (tracks with corrected gain beyond +25 dB are nevertheless very rare). The encoded material would exhibit strong artifacts with ReplayGan set with Track Mode (they won’t be audible otherwise, except maybe as a subtle form of distortion – it could explain some recent complains about musepack and classical music). With RG enabled, even untrained people will be shocked by the terrible ringing that run across this musical material. MPC, with --standard profile, and to some degree --extreme and also –insane is apparently not sensitive enough to handle low volume situation.


At this stage of my account, some people would be probably tempted to claim that such issue is normal with perceptual encoding, and that all other formats will suffer from the same issue in this specific playback condition. But a quick comparison would immediately deny all validity to this idea. I’ve compared musepack --standard to comparable MP3, AAC and Vorbis presets, and these competitors showed the ability to encode properly (no ringing, flat lowpass at high level) the same material. Even stranger, MP3 at 128 kbps, or Vorbis at 90 kbps (!), or AAC (faac!) at 100 kbps perform *much* better than musepack --standard. In other words, perceptual encoders (at least modern one) could handle this situation transparently at mid/low bitrate, even with VBR; only musepack fails, and badly. It might be interesting to note that the VBR model is apparently flawed: with --standard, the bitrate drops to unusual value (110…140 kbps), and quality to an even more abnormal threshold. An illustration (graphical – listening tests were performed upstream - click for link) could make things easier to understand:



I’ve also uploaded an additional gallery - the last one looks very weird! and sounds even worse as it looks.


The ringing, and the austere lowpass, are obvious on these screenshots. Quality is objectively worse than MP3@128; subjectively speaking, the audibility is –as usual- linked to various conditions: hardware, player settings (RG or not), listener’ sensitivity to ringing. Some users won’t notice it, some others will be frightened. The important point to note here is that other audio formats have no problems; my purpose wasn’t to make an infertile comparison between MPC and other. Based on this comparison, I’m tempted to say that MPC could rejoin them with some tuning. Anecdotal point: LAME had recently serious issue (which also concern 3.90.3 ABR at mid/low bitrate) and they were recently solved by developers. I think Gabriel worked on an adaptive ATH threshold, and it might be a lead for MPC developer or for some users which are interested to play with current encoder switches.


I’ve uploaded some samples. The gain for short samples is necessary different from the gain of complete sample; but I’ve tried to cut sample with similar gain. The WavPack samples uploaded have all the native gain and the track_peak of the full track. I’ve also duplicated the track gain to the album gain.

http://guruboolez.free.fr/MPC/quiet_tracks_replaygained.zip

Two appendix in this zip file : a piano sample for which track gain for the sample doesn’t really match to the track gain of the full track (+40 dB instead of +25 dB) ; and a very noisy track for which musepack doesn’t have any problem, despite of high gain correction.


This report is probably the last one I’ll do for MPC (a developer have claim their lack of interest for improving classical at --standard), but I nevertheless hope it will help to improve the encoder. Playing with command line (in order to change ATH or noise sensitivity) might be enough to solve or reduce this issue; therefore, every MPC user could contribute. In the meantime, users should be aware of this issue.

This post has been edited by guruboolez: Jun 23 2005, 10:26

[Gabriel]

Isn't the purpose of track gain to be able to use the tracks in a compilation of tracks from different albums?
In this case, why would track gain be dependant of the gain of its neighbours tracks in the original album? If someone wants to listen a full album, he would use album gain, not track gain.

[Frank Klemm]

Isn't the purpose of track gain to be able to use the tracks in a compilation of tracks from different albums?
In this case, why would track gain be dependant of the gain of its neighbours tracks in the original album? If someone wants to listen a full album, he would use album gain, not track gain.

This is not an article about Musepack, but about flaws in the current ReplayGain
concept and implementation. Low level artefacts are related to this problem,
but this article is about ReplayGain and problems with ReplayGain. ReplayGain
ist *not* perfect at all, there's a long list of problems. Actually I use only
album based ReplayGain, other modes have a lot of problems which may significantly
reduce enjoy music.

Okay?

--------------------------------------------------------------------------------------------------------------

We should (always) start from the user application side.
And there there are at least three scenarios.

1. You want to listen to an album in original order. You want to remove volume differences between albums. The aim is to avoid volume differences between albums with different reference level, which is typical for album mastered in different decades. The effect is like a computer's turn on the volume control between albums (and by the way can also implemented in this way when the computer has access to the volume control via RS-232, RS-422, Cable LAN, Wireless LAN, IR-RC, IR-DA, USB, IEEE-1394, CAN, ...).

2. You want to listen to an album in original order. You want to remove volume differences between titels. The aim ist to reduce loudness differences inside an album, maybe also inside longer or very dynamic tracks. These loudness adjustments must be inaudible at all. No clicks, no distortions, no incredible noise boosting, no obviously loudness changes.

3. You want to shuffle titles of different album and you want to remove loudness differences between the titles. The goal is again: No clicks, no distortions, no incredible noise boosting, no obviously loudness changes.

Okay? Always start from the user application side, never start from the SSE and 3DNow!
assembler code side! Even when you like assembly programming on Intel.
Okay?

--------------------------------------------------------------------------------------------------------------

Now from the programmers geek side.

Solution for 1 is the album based replaygain and this work really great.

Solution for 2 is the proposal above. There should be a free parameter (with a useful default value) which controls the speed of the AGC. Someone want only slight loudness correction especially on title borders, another people wants to remove nearly all dynamic.

For album with silence at the title boundaries title based replaygain often works.
When there's only little loudness difference inside the album also the album gain works (see examples 2)!
There are some conceptional flaws and some flaws in the implementation.
* From time to time there are titles with (psychological) much too much title based replaygain gain
** short titles
** silent titles which must be silent at all
* file boundaries not at the real title boundaries
* live albums / classic without gaps
* noisy albums with little, but audible noise between tracks
* albums with DC
When an album has one of these problems, title based replaygain works poorly.
(see also Examples 1).

Solutions for 3 are also not so easy to implement. First of all we have problems with
boundaries (shifted, no silence). But it is not the task of replaygain to find useful
cuts or to fade between live titles during playback.

Primary goal is to remove loudness differences. Secondary goal is to avoid boosting
of intentionally silent titles. Third goal is to avoid noise boosting.

Short and very silent titles should not be boosted as much as replaygain's function GetTitleGain() says. This is important to avoid boosting of short and intentionally
silent titles.

To avoid noise boosting, ReplayGain's title gain must be limited. ReplayGain's album gain makes less problems, I only found albums with gains between -12.3 dB and +9,4 dB.
Within the title gain I found gain from -13,7 dB to +30,9 dB.

--------------------------------------------------------------------------------------------------------------

Even for 2 AND for 3 the current ReplayGain title gain is a really good implementation.
There are some flaws affecting the usage of ReplayGain title gain for problem 2 AND for problem 3. These should be fixed. In combination with a solution of the cut problem then ReplayGain title gain is also a great solution for problem 3.

Problem 2 must be solved completely different. Even when the ReplayGain title gain problems are solved, there are enough problems remaining when you want to use it
with music with
* file boundaries not at the real title boundaries
* live albums / classic without gaps
* noisy albums with little, but audible noise between tracks
* albums with DC

Examples: see next posting(s)

This post has been edited by Frank Klemm: Jul 9 2005, 15:23

[guruboolez]

This is not an article about Musepack, but about flaws in the current ReplayGain
concept and implementation. Low level artefacts are related to this problem,
but this article is about ReplayGain and problems with ReplayGain.

I beg you pardon, but you've apparently miss an important point: the problem occurs with ReplayGain and Musepack.
ReplayGain + Lame = no problem
ReplayGain + Vorbis = no problem.
ReplayGain + AAC = no problem (checked recently with Nero and previously faac)
ReplayGain + DualStream = no problem
ReplayGain + WavPack lossy = no problem
ReplayGain + MPC = problems

I won't say that ReplayGain is the cause of the problem. It's obviously MPC which can't be used safely in some situation at some encoding profiles.

ReplayGain ist *not* perfect at all, there's a long list of problems. Actually I use only
album based ReplayGain, other modes have a lot of problems which may significantly
reduce enjoy music.

Okay?

What should I conclude? That a problem shouldn't be considered as relevant because you have other listening habits? That's fine as long as you consider MPC as your own tool, dedicated to your own purpose and to people sharing the same behavior, and if you don't plan to improve the obvious bug occuring in other situations. But don't wonder if musepack will quickly loose its audience with such reasoning.

This post has been edited by Dibrom: Jul 11 2005, 00:14

[Dibrom]

I won't say that ReplayGain is the cause of the problem. It's obviously MPC which can't be used safely in some situation at some encoding profiles.

I think in fact it is more correct to say that MPC can't be expected to provide good results in all situations where postprocessing of the encoded file might occur, unless the user intervenes at encode time and provides the encoder with some information that can correct for this after the fact.

This still shouldn't come as a big surprise for a lossy codec.

Yes, other codecs may not exhibit this problem, but at least in the case of some (e.g., the adaptive ATH Gabriel mentioned), it appears that they rely on tradeoff's as far as compression efficiency is concerned here. And furthermore, without 2-pass encoding being used, such an adaptive method is going to be much more imprecise.

What should I conclude? That a problem shouldn't be considered as relevant because you have other listening habits?

The inverse can be said just as well. Should we conclude that MPC is buggy because it doesn't perform as well as you'd like after you modified the encoded file, invalidating the assumptions the encoder made about masking, etc.?

That's fine as long as you consider MPC as your own tool, dedicated to your own purpose and to people sharing the same behavior, and if you don't plan to improve the obvious bug occuring in other situations. But don't wonder if musepack will quickly loose its audience with such reasoning.

Calling this a bug seems like a real stretch to me. It's not a problem with the encoder itself, it's a problem that arises when using the encoded file in certain situations that the encoder doesn't know about. Furthermore, it would seem definitely not to be a bug since the problem goes away if you in fact give the encoder the correct information at encode time (e.g., ath adjustment).

I agree with Frank -- this is not a problem with MPC but a problem with using Replaygain in this way. And that doesn't mean that there cannot be some sort of solution found, but I don't think that implementing some sort of imprecise changes in the psymodel to compensate for this special case -- sacrificing efficiency along the way -- is the right answer.

I think the proper solution is to implement some sort of multi-pass encoding system that allows for the encoder to interact with Replaygain more intelligently, if this is the way you plan to use your encoded files. Maybe this would be done by calculating the Replaygain values for the file before encoding, and then using the track gain as an indicator for ATH adjustment and the like. Of course this won't be an automatic process (you would have to specify some flag like --2pass), but I don't think it should be -- for people not planning to use their files this way, it should be unnecessary for them to have to make the efficiency sacrifices resulting in larger files.

Note: Sorry I accidently hit edit on your post instead of reply

[Gambit]

I think in fact it is more correct to say that MPC can't be expected to provide good results in all situations where postprocessing of the encoded file might occur, unless the user intervenes at encode time and provides the encoder with some information that can correct for this after the fact.

Erm, we are talking about Replaygain here. In the end, it's the same as adjusting the volume knob. I would hardly call that "postprocessing".

[Dibrom]

I think in fact it is more correct to say that MPC can't be expected to provide good results in all situations where postprocessing of the encoded file might occur, unless the user intervenes at encode time and provides the encoder with some information that can correct for this after the fact.

Erm, we are talking about Replaygain here. In the end, it's the same as adjusting the volume knob. I would hardly call that "postprocessing".

Well call it what you like, but when you make significant changes to the playback level (e.g., what is happening here with classical music), it's just about as good as postprocessing as far as a psychoacoustic encoder is concerned.

The file is being processed (volume changed significantly) in a way the encoder does not expect, after the file has been encoded.

[rjamorim]

Well call it what you like, but when you make significant changes to the playback level (e.g., what is happening here with classical music), it's just about as good as postprocessing as far as a psychoacoustic encoder is concerned.

The file is being processed (volume changed significantly) in a way the encoder does not expect, after the file has been encoded.

But if something as simple as turning volume knobs is tripping the codec, then the problem can't be in the volume knob :B

[Dibrom]

Well call it what you like, but when you make significant changes to the playback level (e.g., what is happening here with classical music), it's just about as good as postprocessing as far as a psychoacoustic encoder is concerned.

The file is being processed (volume changed significantly) in a way the encoder does not expect, after the file has been encoded.

But if something as simple as turning volume knobs is tripping the codec, then the problem can't be in the volume knob :B

Umm.. It's not "something as simple as turning the volume knob" that is "tripping the codec." If it were, this problem would be much more pervasive and would occur practically everywhere, instead of being limited to the rather specific conditions (i.e., significant boosting of highly dynamic music to levels that invalidate encode time assumptions about masking, etc.) that guruboolez has brought up.

[rjamorim]

Umm.. It's not "something as simple as turning the volume knob" that is "tripping the codec."  If it were, this problem would be much more pervasive and would occur practically everywhere, instead of being limited to the rather specific conditions (i.e., significant boosting of highly dynamic music to levels that invalidate encode time assumptions about masking, etc.) that guruboolez has brought up.

Well, since it doesn't shows up at all in other codecs, maybe it is the encoder we're discussing here that is making wrong "encode time assumptions about masking, etc."?

After all, if other codecs' psymodels can handle this issue (without two passes!), why can't Musepack's? That is what I am failing to understand. Granted it's a rare ocurrance only reported by Guru so far, but still, there's the point that no other codec suffers from it - even little tunes ones like Faac.

This post has been edited by rjamorim: Jul 11 2005, 00:44

[Dibrom]

Well, since it doesn't shows up at all in other codecs, maybe it is the encoder we're discussing here that is making wrong "encode time assumptions about masking, etc."?

After all, if other codecs' psymodels can handle this issue (without two passes!), why can't Musepack's? That is what I am failing to understand. Granted it's a rare ocurrance only reported by Guru so far, but still, there's the point that no other codec suffers from it - even little tunes ones like Faac.

There are two ways I see that this could be handled, and is probably what is going on in the other codecs:

1) Adaptive ATH like what Gabriel was talking about. It should be pretty obvious why a two pass solution is going to be better than this. The encoder has to adjust the ath according to some sort of running average, and this is something that will have to be hand tuned probably. If you adjust too fast, you lose a lot of efficiency. If you adjust too slow, the user will hear artifacts that you were trying to adjust to prevent. How do you suppose you are going to decide this? There's not going to be a perfect way because you're working with incomplete data the whole time unless you go 2-pass.

2) Simply encode with a less aggressive ath level. This is probably what is happening more of the time. The problem here should be pretty obvious too -- you're always less efficient than you could be. Maybe this isn't a problem if most people are happy with the bitrate though.

Maybe something else is going on and I'm simply missing it, but from the sound of what has been said, I don't think this is the case.

[rjamorim]

Thanks for your clarification.

So, from that, one can infer that Musepack is being a little too conservative on choosing an encoding parameter (ATH) to favor on the bitrate efficiency side?

[Dibrom]

Thanks for your clarification.

So, from that, one can infer that Musepack is being a little too conservative on choosing an encoding parameter (ATH) to favor on the bitrate efficiency side?

This is what I suspect probably, yes.

'Conservative' here could actually mean 'more exact,' since things seem to sound fine until you push them into the situation guruboolez is describing. Unfortunately, when that happens there's no headroom anymore and you sometimes might hear artifacts.

This post has been edited by Dibrom: Jul 11 2005, 01:07

[CiTay]

'Conservative' here could actually mean 'more exact,' since things seem to sound fine until you push them into the situation guruboolez is describing.  Unfortunately, when that happens there's no headroom anymore and you sometimes might hear artifacts.

Yes, that was my feeling too. With that in mind, i think we should look at the possible fixes, i.e. --ath_gain -14 / Ltq_offset tweaks, and see if we can work something out from there. If it's just an increase of 3, 4 kbit (and it would only be needed for --quality 5/--standard, from what i understand), it could easily be added in a new release. After all, the bits that were shaven off in 1.06~1.1 would simply be added again for one profile.

[guruboolez]

With that in mind, i think we should look at the possible fixes, i.e. --ath_gain -14 / Ltq_offset tweaks, and see if we can work something out from there. If it's just an increase of 3, 4 kbit (and it would only be needed for --quality 5/--standard, from what i understand), it could easily be added in a new release. After all, the bits that were shaven off in 1.06~1.1 would simply be added again for one profile.

I fear that this kind of fix can't be proposed as default. The increase in bitrate would probably appear as excessive. Don't forget that the current 1.15 series is not on par with previous MPC encoders: bitrate is systematically higher, for an identical quality compared to 1.14 on a wide majority of samples. Some users have already complaint about it. 1.06 was the most "efficient" encoder; but when 1.1 was released, the bitrate was highered again to reach the same level than pre-1.06 encoders.

I've tried to evaluate the impact of the --ath_gain -14 on standard preset with my 150 short samples (edit: classical ones - it might differ with other music). It should give an accurate idea of the inflation:

- 1.01j --standard = 185 kbps
- 1.14 --standard = 188 kbps
- 1.15v --standard = 196 kbps
- 1.15v --standard --ath_gain -14 = 214 kbps

As you can see, 1.15v without additional switch already increase the bitrate compared to the latest beta. The inflation could reach 15 kbps per album (tried with harpsichord music) without experiencing any gain in quality. And compared to 1.01j, the difference is even higher - and also recall that I've tested 1.01j to sound better on a panel of 15 samples.

Now the ath_gain command would probably increase the quality, but the price to pay is a bitrate, with --standard profile and 1.15v, comparable to older encoder at --extreme profile.

The html files created with MisterQuestionMan:

- mppenc 1.01j
- mppenc 1.14
- mppenc 1.15v
- mppenc 1.15v --ath_gain -14

This post has been edited by guruboolez: Jul 11 2005, 11:56