EDIT 02.27.2004 - Here is the sample : http://www.hydrogenaudio.org/forums/index....ST&f=35&t=19099

Hello,
First let me say that I'm a complete newbie in MPC (didn't encode anything until 2 days ago).
The common opinion in here is that MPC is way better than MP3 at high quality encoding, so I thought I would use --quality 5.

Since I prefer testing than trusting, I took a bunch of tests samples : Amnesia, short, vilbel, spahm, ravebase, and a sample of Rebecca Pidgeon- Granmother (an audiophile recording, not difficult for encoding).

And... big disappointment : amnesia sound UGLY ! Much worse than Lame APS : it rings.

ABX 16/16 at quality 5, 16/16 at quality 6, 8/8 at quality 7.

It still rings at quality 6, but the problem is gone in quality 7, there is just some unimportant noise left (still worse than Lame APS IMHO, but as Lame adds more noise, it becomes a matter of taste)

If we discard Short, that is an artificial sample, I find one big failure for Lame 3.90.2 APS (Badvilbel), and one big failure for Mppenc 1.14 beta --quality 5 (Amnesia) (decoded by Winamp plugin 0.97f).

Note that amnesia is a partly artificial sample, since it was normalized. The original peaked around -6 or -8 db AFAIR. On the other hand, it is a vinyl copy, so if it had been mastered on CD, it might have peaked higher anyway.

Amnesia is available is the test sample page. What do you think of it ?

I also searched for tests about MPC quality. I found Guruboolez one, that concluded that MPC 1.06 added a coloration to the sound of his sample in standard and extreme, but performed way better than Ogg and AAC. http://forum.hardware.fr/forum2.php3?post=...&trash=&subcat=

I'm not able to elaborate a serious discussion upon this purpose but ..

I'm coming very suspicious on real "auditive" quality of mpc encoding file ...
I'd many times a bad audio feeling on High freq (like saturation, glitches, lack of dynamic, choked, crushed ??)

i'm not confortable to explain that in english

I test a Mpc and Ape coding and notice that these issues doesn't append with Ape .. With Ape sound is clear on hig Freq without saturation and correct dynamicity

I'm sure that it's difficult to criticize Mpc that appears be adopted by expert community, so i think this post before .. and if i am not able to prove but that i can demonstrate and reproduce for myself, i'm sure of my feelings

Mhh, do you mean Ape "Monkey's audio", or Ape "Alt Preset Extreme" ?

Sorry to tell this since you are on my side, but we need blind tests, only blind tests ! When you get the feeling, decode the file to wav, cut it into a short sample, and run PC ABX.

EDIT: Jump over this message, the below is not the problem in question:


I wouldn't draw more conclusions about ringing that what there really is..

Amnesia is a sample which has high frequency "ringing" with MPC standard even with higher quality levels depending on hearing. High frequency ringing is a specfic problem, which is quite common for all codecs in some cases, but which is not unambiguous.

Here's a small image of MPC standard amnesia:

You can see that there are small spikes there, response is not totally smooth, like it is with Dibrom's APS. So, you may ask, does this apply for other cases as well, meaning that if the response is not always totally smooth, it will always "ring". The answer is no. It's very common with all codecs that there are high frequency spikes for example with hihat. The issue here, why people with good high frequency hearing may hear ringing artifacts in this case, is probably, because the sound is otherwise quite homogenic, but the small spikes break this reasonably homogenic sound. Of course, if there was a hihat hit there creating a big high frequency spike, nobody would notice this type of artifacting. But as I said, ringing is not unambiguous, and really depends on the case.

Is high frequency ringing generally a problem for MPC? In my opinion not a very big problem. But sure, it should perform better with this clip. I think one problem here was, that amnesia is a very good sample to illustrate this problem (and it was also used by Dibrom when tweaking APS), but it's a 48kHz sample, and MP+/mpc didn't support 48kHz until sometime after Klemm took over iirc (and of course people are too lazy to downsample), so I think it has kinda been forgotten in mpc testing, maybe. I agree that this issue should be further examined.

Can you make conclusions, that because there's riniging with amnesia, there's "saturation, other glitches, lack of dynamic, choked crushed high freqs". No, you can't. You better provide some proofs for these claims.

I doubt that peaks above 18 kHz, has any effect on what I hear : I can't hear anything past 16 kHz !

But I forgot that it was a 48 kHz sample, I'll downsample and retest, then I'll go hunting for similar samples... (say "bzz" !)

EDIT : Uh ? No, it is 44.1 kHz....
EDIT : oops : I'm working on a downsampled version from the begining

I haven't launched ABC/HR yet. But the sound of this amnesia sample seems to be close to the Jump sample (the short one) I provide some weeks ago. But the artifact was something like a smeared sound.

There is maybe a good way to know if the adapatative lowpass of mpc responsible of a audible degradation is : encode amnesia with the old mppenc (< 1.01j) at --insane, which never lowpassed anything.

I haven't abx it. I focuse on seconds 00.000 - 01.500. Just noticed two kind of obvious flaws.
* mp3 blur the sound : smeared details on the whole passage
* mpc is sharp enough, but I hear something wrong at ~0:500, like an excessive contrast.

I opened the mpc file in CoolEdit ; here are the results :

mpc representation
original représentation

I suppose that the big empty spaces are directly responsible of the « contrast » I heard in ABC/HR. I checked other parts of the mpc file : generally, the visual degradation is not as pronounced as I saw it in this 0.500 area. However, some passages I didn't test (above the first second) are badly degraded too, and I suppose that they are ABXable too.
Maybe the explanation of the disappointment or surprise of Pio2001.

I understand (and share) the feeling of Pio2001. The last seconds are buggy, like colorated by an small electric noise (I suppose that the technical word for this is « HF ringing »). This is exactly the same feeling I had for the Jump sample, maybe stronger for amnesia. I began to heard it when I tested the standard preset with the old 1.01j codec : the whole file, from the beginning to the end, is audibly different. The 1.14 codec is a real progress : ringing seems (for me) to be reduced, and really annoying at the second part of the file.


part of amnesia : 4.3-5.9
mpc version : 1.14 --standard --xlevel
mp3 version : lame 3.90.2 --alt-preset standard

Yeah, no excuse for me (except too little time) . The main problem is not that what I was presenting, rather the problem here is obviously dropouts, like guruboolez pointed out.
What surprised me is that even with old mp+ encoder and using VBRmode2 and minSMR 1-3 didn't still give as solid spectra as APS with this clip. Very weird.

Here's my Amnesi5.mpc file (you'll maybe have to right click and save target as).

Could you check that it has no more problem than yours ? And upload yours for me to check ?

In the meantime, I checked 26 tracks featuring transients from my CDs without finding any flaw in MPC encoding yet. The hunting goes on...

EDIT : did you notice that Lame APS gives 317 kbps ?

Didn't ABX it.
I just quickly played A-B A-B, and only one file was obviously and immediatly different : yours ! What did you make ?


EDIT : my encoding is here (right click/save target)
EDIT2 : link fixed

ABX 8/8

Sounds like hearing the spinning blades of a helicopter. I was using the 48kHz version directly without downsampling and mppenc 1.14 with no additional switches. Those sure are some serious dropouts, something I haven't witnessed before with mpc.

There are some samples which make mpc trip up bad, but those are mostly isolated cases. Mpc doesn't seem to have a specific type of sound it can't handle though, like transients for example.

edit: Pio, I just listened to your sample and I think it sounds different than what comes out of "my" mppenc. I believe it has some more noise in the stereo image. Stand by for abx

edit 2: There is some noisy garble which, when listened to with headphones, seems more "outside" than the rest of the sound at ca 0.4 secs. abx 8/8.

edit 3: wtf?! I just listend to your sample, guru, and it also sounds different, than Pio's and mine. I think it sounds worse. What's happening?!

I can ABX 0.0-1.5 every time; the MPC file sounds 'garbled,' or also as if the distorted lead didn't have any reverb on it in the MPC version but did have it on the source file. I haven't compared it to APS yet.

I have to say that's a very, very particular situation. If the sound wasn't as repetitive, I'd probably only be able to catch the difference half of the time. I'm also using speakers instead of headphones, so I'm not sure what kind of a difference that makes.

I know of a similar situation on one of my CDs, and am going to dig it up now to see if it exhibits the same problem.

EDIT: "every time" meaning 15/15

Yes, exaxctly !

Guru, here's my original : Right click and save target as

I think your MPC file sounds basically the same as mine...

I tested it, and I'm not totally agree. The degradation, at the end of the file, is the most awfull I never heard with a musepack encoding, at --standard. The amnesia sample I downloaded some month ago on HA server is not so critical for mpc.

Here are ABC/HR results (quick test) :





I rated 1.5/5 the mpc sounding, at the last second. According to my imagination, sound is not like an helicopter, but more trivially like the oil bubbling on a electric fryer :-þ

I would like to refer to this same link ... from gURuBoOleZZ
http://forum.hardware.fr/forum2.php3?post=...&trash=&subcat=

that i read after my last post

and found a similar evaluation referenced as "Piou piou .." for AAc psytel encoder (i have'nt test it) but also for "Musepack SV7 (mppenc 1.06) --standard (ou --quality 5)"

I'm sure of my hears feeling ..
but need .. of course to ... produce some proof .. and determine wich freq, instrument, or typical and repetitive sound curves produce this type of bad noises ...

I identify the "PIIIii" of piou as a saturation and distorsion defined as « HF ringing »
and the "IOUUUU" like an rapid audio decompression ..... that i think is a perceptive mechanism of brain treatment
"HF" is a bad usage of course .... it's probably not 16khz ...


My opinion is that "noises" are, for me, very difficult to support and produce, because i identify this as a repetitive situation with Mpc audio file, a suscipicious for that codec ...
Maybe, i hope, there is some solution to maybe fix or probably prevent that !!

I have also a real problem with QCD ... Audio player .. .. mpc file are insufferable to hear ... do you know this issue ??

I know some comparative tools and sites like ... PcABX and well-know the "blind-test" process coming from Pharmaceutical process of usa gov agreement .... and the different "sound test series" like Fatboy, Castanet, MpegData, Xiph and some good site like FF123, mp3.radified, Buschmann & Klemm, rareties etc ... others ..???

Or And will be a good first process to test .. audio devices and environement before ...
- ourself hear evaluation (ok include in PcAbx site..)
- some special audio situation .. http://www.dogstar.dantimax.dk/testwavs/ & http://www.dogstar.dantimax.dk/testwavs/
- Devices (for me Audigy I - ATP5 - )
- Softs for me ...... Win XP Pro - Foobar ...

So, is up to me to do the tests ... i never do that ...so

where can i Found the good procees method - tools and audio reference files
Here is maybe the good process steps??
1) evaluate your ears
2) test & verify your material & software
3) find a good reference set of audio file for testing (including freq, audio signal config, voice and music, etc ..)
4) find the correct blind-test soft and scope & visualization of audio signal ... with an acceptable protocol
5) make the test ... verify condition ...
6) publish (wich form ..?)

Is there anybody produce a such protocol & process description?

Hi,

I moved some changes in the SV8 tree back to SV7, hopefully this will reduce some of the problems.
I can't move all SV8 changes to SV7, the format do not support some of the new features.

Bitrate is increased by typically 3.1% compared with 1.14. Most of this bitrate increase is related to
some modificitions in the tonality estimation model (which was introduced with 1.15g).

There are also some other changes which have nothing to do with quality (some counters are now
64 bit to overcome problems with very long movie sound tracks), and these must be tested very carefully.

mppenc 1.15q for Linux will be uploaded in the next minutes. Windows version will take a little bit longer.

amensia encoded with Frank's new encoder:

http://sjeng.org/ftp/vorbis/amnesia.std.mpc
http://sjeng.org/ftp/vorbis/amnesia.xtr.mpc

I'm an idiot : I used 1.95i codec, and not 1.15q

Thank you, Garf.

I tried to compare 1.14 and this new 1.15q, at --standard.
Sound is really close, without real improvement at the beginning (dropout at ~0.500), and seems to be identical at the end (helicopter sounding).






With some doubts, I found 1.15q a bit worse at the beginning as 1.14. Before the dropout, the sound of 1.15q seems to be sour ; I hear it on 1.14 only after dropout.

I'm an idiot : I used 1.95i codec, and not 1.15q

I compared directly 1.14 & 1.15q with an ABX test, at the beginning of the file (0.0-0.5).
One file was corruped by a sour, metallic and unpleasant coloration (the same I can hear, amplified, at the end of the file). I had some difficulties to perform a concluant ABX session. After three trials, the last one, with concentration :



13/16, in favour of... 1.15q, more natural, without this special coloration at the beginning.

Just notice an epistemologic problem
My amnesia.flac (downloaded on HA server some weeks ago) is 44100 Hz
Pio2001 amnesia sample is 44100 Hz
Garf amnesia is 48000 Hz.

Now, I opposed a 44100 Hz encoding (my 1.14 one) to a 48000 Hz one (provided by Garf).

The original I used is at:

http://sjeng.org/ftp/vorbis/amnesia.flac

I'm stupid : I used 1.95i codec, and not 1.15q

Sorry...

Perform again the test, after a big, black coffee.
First comment : the degradation is, with the original sample provided by garf, less annoying, with both codec. Consequently, there are more difficult to hear, especially at the beginning. The end of the file is really better now


BEGINNING :

1.14 : small coloration (saur sound), and the dropout at the same place
1.15q : small dropout, no coloration




ENDING :

1.14 : coloration but not pronouced. Unstable (trembling) signal
1.15 : nothing wrong : I wasn't able to distinguish the mpc file before the ABX test.




Quality is in progress ; bitrate too :

1.14 : 225 kbps // 260 kbps
1.15q : 270 kbps // 311 kbps

For the record, the sample I've been using is the same as the one Garf provided (only mine is a tad longer but otherwise bit identical). I double checked, I really am using 1.14 beta at quality 5.

Garf's 1.15q std sample is still ABXable vs Garf's original (8/8).
I can't abx my 1.14beta std encode vs Garf's 1.15q std encode.
I can however abx Pio's 1.14 std encode vs Garf's 1.15q std encode.

edit: damn, when you hit tab and then space, your post is submitted...

I took more than 200 CDs and asked myself which ones would maybe have transients the would be close to Amnesia, and cause problems with MPC.
I found 41 tracks with more or less transients, but none of them have as strong ones as amnesia, and none of them caused obvious problems to MPC encoding (I didn't search for artifacts smaller than the ones in Amnesia, and I only listened to the parts with transients), which is comforting :

Autechre -- Chiastic Slide -- 03 Tewe.wav
Unknown Artist -- Unknown Title -- 08 Track08.wav
Arcana -- Cantar de Procella -- 01 The Opening of the Wound.wav
Kraftwerk Ralf und Florian Kristallo.wav
ELEGIA -- SOUND WITHIN -- 06 rainfall....wav
C.J. Bolland -- Electronic Highway -- 07 Spoof.wav
Autechre Avil vapre Track01.wav
Autechre Anvil VapreTrack04.wav
Unknown Artist -- Unknown Title -- 09 Track09.wav
Front 242 -- Angels Versus Animals -- 02 Angel (Wipe Out).wav
Front 242 -- 06-21-03-11 Up Evil -- 13 Religion (Pussy Whipped mix).wav
Lenny Dee & Ralphie Dee - I Control Your Body.wav
Tellurian - Speed of Light.wav
Annihilator - S.D.R (Bam Bam).wav
Front 242 -- Angels Versus Animals -- 07 Der Verfluchte Engel.wav
Electric Universe -- Stardiver -- 08 Astral Voyage (Edit).wav
Prototype 909 - understand.wav
Electric Universe -- Stardiver -- 11 Alien Encounter (Part 1).wav
Unkknown Artist -- Unknown Title -- 08 Clavecin.wav
Air Liquide - sun progress.wav
The Irresistible Force - space is the place.wav
Mimir -- Mimir -- 06 December 3. 1989.wav
Mimir -- Mimir -- 07 December 2. 1989.wav
Electric Universe -- One Love -- 08 Visiting Venus.wav
Kldnge - No Future.wav
The Future Sound of London -- Omnipresence.wav
Beaumont Hannant -- Teqtonik.wav
Autechre -- Doctrine.wav
The Chekists -- Global (remix).wav
Sun Electric -- Sarotti.wav
AFX (aka Aphex Twin) -- .215061 (Analogue Bubblebath 3).wav
Difficult Child -- Fear.wav
Age -- Eine fremde Lebensform.wav
Takashi Harada -- Palme Symphony -- 03 Koram … sept temps.wav
Takashi Harada -- Palme Symphony -- 09 Pu et Mu.wav
Takashi Harada -- Palme Symphony -- 11 Sur un bateau.wav
Takashi Harada -- Palme Symphony -- 20 Depuis les t‚nŠbres lointaines.wav
Takashi Harada -- Palme Symphony -- 22 Nous y avons cru.wav
Depeche Mode -- Some Great Reward -- 03 People are people.wav
Depeche Mode -- Some Great Reward -- 07 Master and servant.wav
Depeche Mode -- Some Great Reward -- 08 If you want.wav

Hmm, I'm testing the new mppenc 1.15r.

Funny thing is that if the source amnesia.wav is 48kHz, standard sounds clearly better than if the source amnesia.wav is first downsampled to 44.1kHz and then encoded with 1.15r standard.

I can clearly ABA (4/4) the difference between the MPCs depending whether the source has been 48khz or 44.1khz. 44.1khz version sounds clearly worse: noise spreading around and more watery sound. I hear it better if I try to forget the "choir" and don't use too high volume.

I uploaded 44.1kHz here: amnesia44.flac. It's downsampled from the original using CoolEdit2pro quality 999.
Here is the original 48khz amnesia.flac.

edit. Well, you can see also from the spectral view, that if the source was 44.1kHz, the resulting MPC's spectra has larger dropouts.

Where can I see the spectral plots? (44.1 + 48)
Is the noise a continuous noise or is it only sometimes audible?
What say 1.95z6 with 44.1/48/64 kHz?
What headphones / loudspeakers do you use (I hope nothing like HD570/HD575/HD590)?
RMS level in Volts on loudspeakers?
Is there a need of a switch which headphone/loudspeaker do you use for listening?

I have currently some trouble with the email/WWW access.

in the 0.500 sec. area.

mpc 1.14 std representation
original représentation

I heard it with my headphones (both Philips HP910 & BeyerDynamic DT-531)

Below are the graphs for comparing 48khz and 44khz amnesia using mppenc 1.15r standard.

48kHz version 0.43-0.76s sounds closer to the original. (downsampled after encoding to 44khz for easier graphical comparison)
44kHz version 0.43-0.76s has 2 increased spreading noise bursts, probably where the big dropouts happen around 0.53s and 0.64s.

I think this is the most audible difference, but there's all the time positions where 44kHz version gives a bit louder noise bursts than 48khz. I'm not saying that 48khz is transparent, but it's better than 44khz.

Audible with all the headphones I tested.

I'll tell you when I get win32 binaries, or somebody encodes amensia with it.
Case currently has only 1.95e. And I can't access your page atm.

1.14 is uninteresting. 1.15q/r and 1.95z4/5/6/7/8 handles such audio quite different.

Sorry.

Here is the similar portion, encoded with mppenc 1.15r. Dropout is visually reduced. I can't ABX it for the moment.

mpc 1.14 std representation
mpc 1.15r std representation
original représentation

If you try the 44kHz version, you'll see big dropouts again, and should be able to ABX.
amnesia44.flac

ABX result :



Dropout audible with 1.15r ; but sound distorsion is reduced (44K version).

EDIT : 1.15r challenger is the 1.14 version

lowpass the original and decoded signal at

- 6,5 kHz
- 11 kHz

and try to compare originla vs. encoded/decoded.

Mppenc 1.95z6 for Win32 and 1.95z2 for Linux are available.

IMO, 1.95z6 is worse than 1.15r. Dropout, at the beginning, increased again. Confirmed with CoolEdit :

mpc 1.95z6

Lowpassed the original&decoded (mppenc 1.15r standard) using 6.5khz lowpass.

Original vs amnesia44khz.mpc after lowpass 6.5khz
ABA 7/7 propability of guessing <0.1%

amnesia48khz.mpc vs amnesia44khz.mpc after lowpass 6.5khz
ABA 9/11 propability of guessing 0.1%

Anyway, these were definitely very much harder than without lowpass, because most of the revealing spreaded "buzz" burst noise in the 44khz mpc's problem section was gone. But still the "buzz" bursts left were slightly more "monoish" with the original and 48khz version which allowed to distinguishing the 44khz.

Bitrate of 1.15r and 1.95z6?

Also check other samples like Jump.wav.
It doesn't help when
- only amnesia is enhanced
- bitrate increases for all music pieces, also for those without any problems

[quote=Frank Klemm,Feb 23 2003 - 09:11 PM][QUOTE=guruboolez,Feb 23 2003 - 08:34 PM]Bitrate of 1.15r and 1.95z6?[/quote]
I'm really confuse ; I tested 1.95z6 --standard (the worse) against 1.15r --insane. 240 kb versus 166 kb.
I'm sorry (again : this is the second time it happens).

The good news, is that --insane better than --standard is (it was admitted, but maybe not tested )



I will try jump later, this night. I need to clean my hard disk first

Is there any chance to get your audio collection? Most of the titles and groups are completely
unknown to me (although recommended).

The only interpreter I know is Autechre with the album Amber.

I have a decent amount of music similar to amnesia; could you roughly explain what types of sounds you think are causing the problem, helping me limit my search? Is the problem caused by sharp transients (which I thought would be unlikely given MPC's short time resolution) or distortion or parts with a gapper applied to them, or...?

I tried a track with sharp transients and lots of distortion that yanked 200kbps average out of MPC --standard (and 250kbps out of --alt-preset standard), but was unable to ABX any differences between it and the original when using 1.95z6, which I have to say I'm very impressed about.

hi, i don't think to have as much expertise as the ones writing in this topic
i am just asking a questions about preparing trouble samples to hear artefacts easily: you can try it with sound editor (wablb, kooledit)
normalize the volume of one audio channel to max (100%)
the other channel has its volume gain lowered to say 10% or less
encode the file then the artefacts are really easy to find

is this hard test fair to test the encoders ?

Usually, only real-world music or sounds are used to verify psychoacoustic encoders performance, since it's the real target of them.

I found another sample of the same kind. The effect is the same, but less strong. But this time, it is ripped straight from a CD without resampling nor normalisation. And lame APS performs poorly on it.

ABX MPC 1.14 --quality 5 --xlevel : 8/8
ABX Lame 3.90.2 --alt-preset standard : 8/8

The effect is the same : helicopter sound with MPC (twitwitwitwi), and noise with APS (shfwshhs).

Unfortunately, the sample is 1.6 MB, my old ISP doesn't allow uploading more than 1 MB, and the online account of my new ISP is not working.
Can someone PM me somewhere to upload ?

The sample is Transwave -- The rezwalker (london live remix), 0:00 to 0:26

I didn't expect it to be hard. It came out of 5 tracks I chose to test the audiophileness of MPC on silent CD playback, with the computer turned off :

Idir -- A Vava Inouva
Depeche Mode -- If you want
Rebecca Pidgeon -- Granmother
Transwave -- The Rezwalker (london live remix)
Lisa Gerrard and Pieter Bourke -- Sacrifice

The other ones sound OK.

Wow ! Just listened to Amnesia again... seems my listening is very accurate tonigh... Transwave is far less distorded. Too bad it's time to go to bed

Can you try the 48khz amnesia, which sounds better than the 44khz version. Also, use the latest mppenc 1.15r.

OK, but not tonight, sorry

I'll try to get the time to before the week end

I hope that Frank can come up with his magic to solve this the best possible way. I know he has been in contact with Andree B, and they have a pretty good idea what the problem is. The difficult thing is, how to implement the solution in best possible way.