failure for musepack

Topic: failure for musepack (Read 11994 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.

failure for musepack

2002-12-23 17:29:17

Here is a sample I found, extract of a well-known Van Halen title : Jump. The first seconds are really hard to encode, and I expect mpc to feel at ease with such sample. But the surprise is really big : --standard fail (easy ABX), --extreme fail (easy ABX), --insane too, and the --quality 10 test (express test) gave me a 16/20 score ! In comparison, vorbis 1.0 is more difficult for me to abx at -q 6 (but the bitrate is near 400 kbps).

http://membres.lycos.fr/guruboolez/

Can someone confirm my impressions (headphone needed I suppose) ?

failure for musepack

Reply #1 – 2002-12-23 18:00:04

MPC --standard sounds OK to me - didn't try ABX or anything. Maybe a slightly longer sample would help?

failure for musepack

Reply #2 – 2002-12-23 18:14:04

Quote

MPC --standard sounds OK to me - didn't try ABX or anything. Maybe a slightly longer sample would help?

I'm not sure that a longer sample can help you. I've just upload a longer file.
Note that the first try I did was not concluant. I abruptly heard a small but significant difference (noiser sound, brighter too : can not really explain, and my percepcton are distorded by illness). This difference is not removed by higher bitrate.

failure for musepack

Reply #3 – 2002-12-23 20:02:31

no more test ?! Thanks to Chun-Yu.

I tried again. The first two minutes, I wasn't able to hear any difference with any file. Then, ABX tests were concluant again :

Code: [Select]

ABC/HR Version 0.9b, 30 August 2002
Testname: 

1L = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6(114).wav
2L = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6 (179).wav
3L = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc5(114).wav
4R = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc7(114).wav
5L = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc10.wav
6R = C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc8(114).wav

---------------------------------------
General Comments:
première tentative sur le sample 2 (le 1 étant trop dur)
---------------------------------------
1L File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6(114).wav
1L Rating: 3.7
1L Comment: 
---------------------------------------
2L File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6 (179).wav
2L Rating: 2.5
2L Comment: 
---------------------------------------
3L File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc5(114).wav
3L Rating: 3.0
3L Comment: 
---------------------------------------
4R File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc7(114).wav
4R Rating: 3.3
4R Comment: 
---------------------------------------
5L File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc10.wav
5L Rating: 4.6
5L Comment: 
---------------------------------------
6R File: C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc8(114).wav
6R Rating: 4.2
6R Comment: 
---------------------------------------
ABX Results:
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6(114).wav
    13 out of 16, pval = 0.011
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc6 (179).wav
    10 out of 12, pval = 0.019
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc5(114).wav
    12 out of 12, pval < 0.001
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc7(114).wav
    11 out of 12, pval = 0.003
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc10.wav
    11 out of 16, pval = 0.105
Original vs C:\Ma musique\jump\20. Jump [HDCD Remaster] (2).mpc8(114).wav
    12 out of 12, pval < 0.001

The test I performed was blitz-test, really fast.
The codec was the mppenc 1.14, and one with the old 1.79c at --xtreme profile. I decoded all files with the 1.93 mppdec version. Note that the worst file I founded was suprinsingly the --extreme profile from Andree Buschmann. It's hard for me to to organize into a hierarchy the 6 files. Difference are subtile between mpc files. But I'm sure that there is a real progress between mpc --standard, and mpc --quality 10, and that the extreme profile from the old codec is not as good in this sample than the 1.14 one.

failure for musepack

Reply #4 – 2002-12-23 20:28:30

Yes, I hear it

first winabx round (listening and guessing...) mpc --standard --xlevel (mppenc 1.14b)

21:07:54 1/1 p=50.0%
21:09:00 1/2 p=75.0%
21:09:32 2/3 p=50.0%
21:10:24 3/4 p=31.2%
21:10:41 4/5 p=18.8%
21:10:56 5/6 p=10.9%
21:11:55 6/7 p= 6.2%
21:12:17 7/8 p= 3.5%
21:13:11 8/9 p= 2.0%
21:13:16 test finished

Btw, Guruboolez, I've read here somewhere that you hear a difference with (quote) "2000 year old string instruments" encoded with mpc.
I suspect the same problem with "only 400 year old" baroque strings, but I haven't found a clear abx-able sample yet, do you have one?

-tm

failure for musepack

Reply #5 – 2002-12-23 20:36:15

Quote

Btw, Guruboolez, I've read here somewhere that you hear a difference with (quote) "2000 year old string instruments" encoded with mpc.
I suspect the same problem with "only 400 year old" baroque strings, but I haven't found a clear abx-able sample yet, do you have one?

No. The erhu (2000 year old) sample concern LAME --alt-preset only, and not mpc. I recently found the same problem with a 300 year old flute The only « problem » with mpc with these instrument is the average bitrate, not the quality. Yesterday, I encode a whole harpsichord disc with mpc standard : 1.14 gave me a 220 average bitrate, and 1.15alpha reached 235 kbps ! For low bitrate encoding (under 100 kbps) with these recordings, mpc is one of the worst codec I heard.

But I never heard an artifact or a difference between original with a 'classical' encoding with musepack (> standard). I searched, but never found. I'm sometime able to hear subtile difference with --standard, but score are really fragile, and I can't conclude anything from it.

And thanks for the test, and for the results

failure for musepack

Reply #6 – 2002-12-24 01:48:33

mppdec gets some 'overdrives' (whatever that is? something to do with the fact the file is kinda 'hard limited'?), i used --xlevel in cli as well:

Code: [Select]

mppdec Jump_long.mpc
MPC Decoder  SV7  1.1  3DNow/SSE   (C) 1999-2002 Buschmann/Klemm/Piecha/Wolf

decoding of file 'Jump long.mpc'
         to device /dev/audio (Windows WAVEOUT Audio)

  195.0 kbps,    0:07.49, SV 7.0, Profile 'Standard' (Release 1.1)

     0:07.47 (runtime: 6.48 s  speed: 1.15x)

   18 Overdrives, maximum level 34208, rerun with --scale 0.95784

edit: cant abx with ver 1.1 profile standard
edit2: ok, sounds like at the beginning the right channel wont hit in at the right time, but still cant abx that.

failure for musepack

Reply #7 – 2002-12-24 03:06:33

Quote

mppdec gets some 'overdrives' (whatever that is? something to do with the fact the file is kinda 'hard limited'?), i used --xlevel in cli as well:

Code: [Select]

mppdec Jump_long.mpc
MPC Decoder  SV7  1.1  3DNow/SSE   (C) 1999-2002 Buschmann/Klemm/Piecha/Wolf

decoding of file 'Jump long.mpc'
         to device /dev/audio (Windows WAVEOUT Audio)

  195.0 kbps,    0:07.49, SV 7.0, Profile 'Standard' (Release 1.1)

     0:07.47 (runtime: 6.48 s  speed: 1.15x)

   18 Overdrives, maximum level 34208, rerun with --scale 0.95784

I doesn't have this problem. Tried mppdec -> file & -> device, and no errors.
I haven't perform again a test with --xlevel switch : EAC comparison tool indicate a total identity between mpc --standard and --standard --xlevel.

Thnx for reply

failure for musepack

Reply #8 – 2002-12-24 05:30:34

Quote

Quote
MPC --standard sounds OK to me - didn't try ABX or anything. Maybe a slightly longer sample would help?

I'm not sure that a longer sample can help you. I've just upload a longer file.
Note that the first try I did was not concluant. I abruptly heard a small but significant difference (noiser sound, brighter too : can not really explain, and my percepcton are distorded by illness). This difference is not removed by higher bitrate.

Tested MPC v1.14 --standard, abx 16/24 and 11/16, combined is 27/40 (p-val < 0.02). Hard to ABX, with a vacuum cleaner going in the background!

The MPC version adds a lower "growl" to the sound not present in the original... or maybe it removes the growl, I don't know. It's hard to describe...

I wonder if adding some sound to the beginning of the file will improve quality, like it sometimes does with Lame? Will test by copying clip several times over and ABXing a portion from the middle...

failure for musepack

Reply #9 – 2002-12-24 05:39:30

Have you try anything higher than --standard ?
A small and rare difference with --standard doesn't bother me, but a audible difference at --extreme and above is more annoying

failure for musepack

Reply #10 – 2002-12-24 05:58:02

Quote

Have you try anything higher than --standard ?
A small and rare difference with --standard doesn't bother me, but a audible difference at --extreme and above is more annoying

Standard is right on the limit of my perception. The sample clip's noise is similar to the frequency of my screeching processor fans, which makes it hard to hear.

I keep getting somewhat-significant results with Jump.wav (11/16), and I think I can hear a difference. But I can't clearly hear a difference when I listen to the "looped" sample. To make the looped sample, I copied Jump.wav five times in a row to make a 2.6 second clip, encoded the whole clip, decoded it, and then extracted the third loop of Jump.wav (approximately 1.05 seconds to 1.55 seconds) for testing.

guruboolez, I will try a higher setting on Jump.wav, but I may not be able to provide conclusive results until tomorrow (12 hours for me). Perhaps you can test a "looped" sample to hear if it makes a difference?

failure for musepack

Reply #11 – 2002-12-24 06:38:16

I'm tired too, but confident.
I looped the original file, encoded it at --quality 10 (I'm very confident !), opened the waveform in cooledit, cut the third part, and saved it in .wav.
Then I opened ABC/HR, loaded the original, the looped too, and the previous --quality 10 (based on simple Jump.wav) at the same time.

I wasn't able to hear anything B)

I tried again, but at --quality 7 (--insane with mppenc 1.14).
ABX score are now interesting :
40/60 and 27/30 for the second mysterious file. Is one better than the other ? Why xx/60 ?

The details are instructive :

Code: [Select]

  0 of   1, p = 1.000
  1 of   2, p = 0.750
  1 of   3, p = 0.875
  2 of   4, p = 0.688
  3 of   5, p = 0.500
  4 of   6, p = 0.344
  4 of   7, p = 0.500
  5 of   8, p = 0.363
  5 of   9, p = 0.500
  6 of  10, p = 0.377
  6 of  11, p = 0.500
  7 of  12, p = 0.387
  8 of  13, p = 0.291
  9 of  14, p = 0.212
  9 of  15, p = 0.304
  9 of  16, p = 0.402
  9 of  17, p = 0.500
  9 of  18, p = 0.593
 10 of  19, p = 0.500
 11 of  20, p = 0.412
 11 of  21, p = 0.500
 11 of  22, p = 0.584
 12 of  23, p = 0.500
 13 of  24, p = 0.419
 13 of  25, p = 0.500
 14 of  26, p = 0.423
 15 of  27, p = 0.351
 16 of  28, p = 0.286
 17 of  29, p = 0.229
 18 of  30, p = 0.181
 19 of  31, p = 0.141
 19 of  32, p = 0.189
 20 of  33, p = 0.148
 21 of  34, p = 0.115
 21 of  35, p = 0.155
 22 of  36, p = 0.121
 22 of  37, p = 0.162
 23 of  38, p = 0.128
 23 of  39, p = 0.168
 23 of  40, p = 0.215
 23 of  41, p = 0.266
 24 of  42, p = 0.220
 25 of  43, p = 0.180
 26 of  44, p = 0.146
 27 of  45, p = 0.116
 28 of  46, p = 0.092
 29 of  47, p = 0.072
 30 of  48, p = 0.056
 31 of  49, p = 0.043
 32 of  50, p = 0.032
 33 of  51, p = 0.024
 34 of  52, p = 0.018
 35 of  53, p = 0.013
 35 of  54, p = 0.020
 36 of  55, p = 0.015
 37 of  56, p = 0.011
 37 of  57, p = 0.017
 38 of  58, p = 0.012
 39 of  59, p = 0.009
 40 of  60, p = 0.007

The first twenty are non-significative : 11/20. Then, the following ABX score are slightly better : 12/20 [on 21-40] and 20 last ABX are really good : 17/20.
I need a warm-up, like an athlete

In these conditions, it's difficult to conclude that SAMPLE 1 (40/60) is better than SAMPLE 2 (27/30). Sample 1 is the looped file, and the 'easiest' and second sample is the simple_jump file.
Nevertheless, I had a strong feeling that sample_2 different (and more degraded) was from sample_1. I tried to ABX 1 against 2 : 17/20. Details :

Code: [Select]

  0 of   1, p = 1.000
  1 of   2, p = 0.750
  1 of   3, p = 0.875
  2 of   4, p = 0.688
  2 of   5, p = 0.813
  3 of   6, p = 0.656
  4 of   7, p = 0.500
  5 of   8, p = 0.363
  6 of   9, p = 0.254
  7 of  10, p = 0.172
  8 of  11, p = 0.113
  9 of  12, p = 0.073
 10 of  13, p = 0.046
 11 of  14, p = 0.029
 12 of  15, p = 0.018
 13 of  16, p = 0.011
 14 of  17, p = 0.006
 15 of  18, p = 0.004
 16 of  19, p = 0.002
 17 of  20, p = 0.001

...A perfect 15/15 on the end of the test.

You were right : same audio informations can be encode differently !

failure for musepack

Reply #12 – 2003-01-07 06:28:28

kinda hard to belive only my mppdec is 'biased', has any1 else got to the same 'overdriven' situation? (cos if that is the case you been abxing decoder clipping imho)

edit: i got same overdrives with mppenc 1.14 beta and mppdec 1.93j
edit2: @Tim Mervielde&SometimesWarrior: you forgot to mention how the file was decoded?
edit3: if i ignore the decoder warning, file is actualy clipping (tested in sound forge, clipping just after 6s)

failure for musepack

Reply #13 – 2003-01-07 07:37:46

Quote

kinda hard to belive only my mppdec is 'biased', has any1 else got to the same 'overdriven' situation? (cos if that is the case you been abxing decoder clipping imho)
.../... (tested in sound forge, clipping just after 6s)

The artifact appeared at the first second, not at the sixth

failure for musepack

Reply #14 – 2003-01-07 07:46:52

Quote

Quote
kinda hard to belive only my mppdec is 'biased', has any1 else got to the same 'overdriven' situation? (cos if that is the case you been abxing decoder clipping imho)
.../... (tested in sound forge, clipping just after 6s)

The artifact appeared at the first second, not at the sixth

i guess soundforge & its peakmeter arent really reliable then

Code: [Select]

MPC Decoder  SV7  1.93j  3DNow/SSE   (C) 1999-2002 Buschmann/Klemm/Piecha/Wolf

decoding of file 'Jump long.mpc'
         to device /dev/audio (Windows WAVEOUT Audio)

  193.0 kbps,    0:07.49, SV 7.0, Profile 'Standard' (Beta 1.14)

     0:07.47 (runtime: 6.48 s  speed: 1.15x)

   15 Overdrives, maximum level 33735, rerun with --scale 0.97127

edit: eac wav editor as second opinion, this clipping appears at just around 1st s:

Notice