View Single Post
Old 22 June 2005, 05:56 am   #8
guruboolez
Member
 
guruboolez's Avatar
 
Join Date: Aug 2004
Location: Strasbourg (France)
Posts: 13
Default

Problems occurring at -80 dB shouldnt be audible in any situation. Even with a dynamic compressor which would bring lowest part of the signal to a much higher one, I seriously doubt that audible problems might be revealed to the listeners. I have more than 1000 classical music CDs, and even ultra-quiet parts of wide-dynamic compositions (piano, full orchestra) arent at -80 dB. Its typically -65...-50 dB. There is of course some information at -80 dB, but its usually noise, with very little information coming from instruments.

Nevertheless, musepack has poor performances at mid/high profile (--standard, --extreme and sometimes --insane) with low volume tracks. This problem really becomes audible, and sometimes dramatic, after replaygaining. I have many tracks that are measured by RG at > +20 dB; some of them also reach the +30 dB floor: organ, orchestra and chamber music (especially contemporary one) are periodically implicated. Therefore, if you play these tracks with RG track mode enabled, there will be a terrible ringing. I suppose it could be highly improved, because other encoders are performing *much* better even with low setting: vorbis at q4, faac at q100 and also lame MP3 3.97a at 128 kbps offer all (much) better encoding than mpc --standard and sometimes extreme.

Ive uploaded some encodings to illustrate this problem. Reference file is an organ piece, coming from complete Liszt organ music (5CD) played by Olivier Vernet. The RG value for this single track is +29.47 dB; the RG value for the short sample Ive selected is +40 dB. But Ive manually edited the RG value to match the reference one (+30 dB).

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

PCM file could be downloaded in the same folder:
http://guruboolez.free.fr/MPC/Liszt_Choral.wv


The ringing, absolutely terrible at q5 after RGing, is still annoying at q6, but clearly less with q7. With q8 profile, ABXing wasnt a real problem, but the annoyance is near zero (reference sound quality is more questionable). Ive ABXed q9, but failed at q10.

Code:
ABC/HR for Java, Version 0.5a, 21 juin 2005
Testname: 

Tester: 

1R = C:\tests\Liszt\Liszt - Choral MPC Q6.wav
2R = C:\tests\Liszt\Liszt - Choral MPC Q10.wav
3L = C:\tests\Liszt\Liszt - Choral MPC Q5.wav
4L = C:\tests\Liszt\Liszt - Choral MPC Q7.wav
5R = C:\tests\Liszt\Liszt - Choral MPC Q8.wav
6R = C:\tests\Liszt\Liszt - Choral MPC Q9.wav

---------------------------------------
General Comments: 
---------------------------------------
1R File: C:\tests\Liszt\Liszt - Choral MPC Q6.wav
1R Rating: 1.5
1R Comment: awful ringing
---------------------------------------
2R File: C:\tests\Liszt\Liszt - Choral MPC Q10.wav
2R Rating: 4.8
2R Comment: Transparent this time. I've tried to focused on background details (organ mechanic), but obtained questionable pval on ABX (11% chance for guessing).
---------------------------------------
3L File: C:\tests\Liszt\Liszt - Choral MPC Q5.wav
3L Rating: 1.0
3L Comment:  below telephone . 0/5 would be more realistic.
---------------------------------------
4L File: C:\tests\Liszt\Liszt - Choral MPC Q7.wav
4L Rating: 3.0
4L Comment: ringing, obvious but much less annoying than 1R and 3L
---------------------------------------
5R File: C:\tests\Liszt\Liszt - Choral MPC Q8.wav
5R Rating: 4.0
5R Comment: unconstant, fluctuant noise
---------------------------------------
6R File: C:\tests\Liszt\Liszt - Choral MPC Q9.wav
6R Rating: 4.5
6R Comment: ABXable, but variations of noise are very subtle
---------------------------------------

ABX Results:
Original vs C:\tests\Liszt\Liszt - Choral MPC Q10.wav
    11 out of 16, pval = 0.105
Original vs C:\tests\Liszt\Liszt - Choral MPC Q7.wav
    8 out of 8, pval = 0.0030
Original vs C:\tests\Liszt\Liszt - Choral MPC Q5.wav
    8 out of 8, pval = 0.0030
Original vs C:\tests\Liszt\Liszt - Choral MPC Q8.wav
    7 out of 8, pval = 0.035
Original vs C:\tests\Liszt\Liszt - Choral MPC Q6.wav
    8 out of 8, pval = 0.0030
Original vs C:\tests\Liszt\Liszt - Choral MPC Q9.wav
    8 out of 8, pval = 0.0030


---- Detailed ABX results ----
Original vs C:\tests\Liszt\Liszt - Choral MPC Q10.wav
Playback Range: 01.041 to 21.638
    9:22:16 PM p 1/1 pval = 0.5
    9:22:24 PM p 2/2 pval = 0.25
    9:22:30 PM f 2/3 pval = 0.5
Playback Range: 06.556 to 08.678
    9:22:46 PM p 3/4 pval = 0.312
    9:22:52 PM p 4/5 pval = 0.187
    9:23:01 PM p 5/6 pval = 0.109
    9:23:05 PM p 6/7 pval = 0.062
    9:23:09 PM p 7/8 pval = 0.035
    9:23:13 PM f 7/9 pval = 0.089
    9:23:18 PM p 8/10 pval = 0.054
    9:23:22 PM f 8/11 pval = 0.113
    9:23:26 PM p 9/12 pval = 0.072
    9:23:30 PM p 10/13 pval = 0.046
    9:23:34 PM p 11/14 pval = 0.028
    9:23:39 PM f 11/15 pval = 0.059
    9:23:47 PM f 11/16 pval = 0.105

Original vs C:\tests\Liszt\Liszt - Choral MPC Q7.wav
Playback Range: 01.041 to 21.638
    9:18:36 PM p 1/1 pval = 0.5
    9:18:39 PM p 2/2 pval = 0.25
    9:18:41 PM p 3/3 pval = 0.125
    9:18:43 PM p 4/4 pval = 0.062
    9:18:45 PM p 5/5 pval = 0.031
    9:18:48 PM p 6/6 pval = 0.015
    9:18:50 PM p 7/7 pval = 0.0070
    9:18:53 PM p 8/8 pval = 0.0030

Original vs C:\tests\Liszt\Liszt - Choral MPC Q5.wav
Playback Range: 01.041 to 21.638
    9:17:54 PM p 1/1 pval = 0.5
    9:17:55 PM p 2/2 pval = 0.25
    9:17:56 PM p 3/3 pval = 0.125
    9:17:58 PM p 4/4 pval = 0.062
    9:17:59 PM p 5/5 pval = 0.031
    9:18:00 PM p 6/6 pval = 0.015
    9:18:07 PM p 7/7 pval = 0.0070
    9:18:09 PM p 8/8 pval = 0.0030

Original vs C:\tests\Liszt\Liszt - Choral MPC Q8.wav
Playback Range: 01.041 to 21.638
    9:19:28 PM f 0/1 pval = 1.0
    9:19:31 PM p 1/2 pval = 0.75
    9:19:33 PM p 2/3 pval = 0.5
    9:19:37 PM p 3/4 pval = 0.312
    9:19:39 PM p 4/5 pval = 0.187
    9:19:42 PM p 5/6 pval = 0.109
    9:19:46 PM p 6/7 pval = 0.062
    9:19:50 PM p 7/8 pval = 0.035

Original vs C:\tests\Liszt\Liszt - Choral MPC Q6.wav
Playback Range: 01.041 to 21.638
    9:17:26 PM p 1/1 pval = 0.5
    9:17:27 PM p 2/2 pval = 0.25
    9:17:29 PM p 3/3 pval = 0.125
    9:17:30 PM p 4/4 pval = 0.062
    9:17:31 PM p 5/5 pval = 0.031
    9:17:33 PM p 6/6 pval = 0.015
    9:17:35 PM p 7/7 pval = 0.0070
    9:17:37 PM p 8/8 pval = 0.0030

Original vs C:\tests\Liszt\Liszt - Choral MPC Q9.wav
Playback Range: 01.041 to 21.638
    9:20:21 PM p 1/1 pval = 0.5
    9:20:31 PM p 2/2 pval = 0.25
    9:20:38 PM p 3/3 pval = 0.125
    9:20:49 PM p 4/4 pval = 0.062
    9:21:00 PM p 5/5 pval = 0.031
    9:21:21 PM p 6/6 pval = 0.015
    9:21:40 PM p 7/7 pval = 0.0070
    9:21:49 PM p 8/8 pval = 0.0030
http://guruboolez.free.fr/MPC/Liszt_...playgained.txt




Of course, this situation is not very common, and the big artefact only appears with RG enabled, and with trackgain only. But performance is definitively poor (compared to other tools), and the bad surprises comes from an exceptional low performance of musepack at -q5, used to be a champion with this encoding profile (it was ranked first after a blind listening test involving 18 classical music samples Ive performed last summer), and which introduces here terrible artefacts that even competitors at 100 kbps could avoid.

Working in this issue (IMO the biggest quality flaw of the format at q5/-q6) might be a direction for further tunings of the current encoder.
guruboolez is offline   Reply With Quote