(slight) Problem sample

2003-08-03 20:38:30

I found a problem sample for lossy encoder, and especially musepack. Disc is Tosca, from Puccini, conducted by Pappano (original soundtrack of the film, Tosca, directed by Benoit Jacquot).

Description of the problem :
I heard some noise, and an unatural transition between the two first notes (0.00 - 0.60 // 0.60 - ...) ; the second part is more noisy, aggressive - in one word : harsh. I can hear the same problem some seconds later. Sound is slightly distorted, badly coloured.

The sample (Flac - 2,62 MB - 27 seconds) :

http://membres.lycos.fr/guruboolez/AUDIO/samples/Tosca

I find the sound close to the Jump and Amnesia samples - other known problems for mpc. I've no difficulties for ABXing it at --standard profile. I didn't tried at --extreme, but at --insane... with success.
I didn't have too much time, and therefore, I didn't perform serious test at high settings. I spent my free time for testing --standard profile with three encoders :
• beta 1.14
• alpha 1.15r
• alpha 1.95z67 (pre-SV8 release)

I tested the sample on two different points, close each others to an acoustic point of view :
• 00.00 - ~02.00
• 04.00 - ~06.00

In order to avoid some approximation (mood, luck...), I rated alternately both parts five time. I expected 1.14 to be the worse, 1.15r much better, and pre-SV8 close to original. Results are totally different :

Code: [Select]

PART_1
       1.14    1.15r   1.95z

#1     4.5     4.0     3.5
#2     4.0     4.5     3.6
#3     4.0     4.5     3.5
#4     3.6     4.4     4.1
#5     4.1     4.5     3.5

PART_2
       1.14    1.15r   1.95z

#1     4.7     4.3     4.2
#2     4.7     4.1     3.4
#3     4.5     3.5     4.0
#4     4.5     4.2     4.0  
#5     4.5     3.6     4.1

On part_1 : 
· 1.15r won four times
· 1.95z lost four times

On part_2 : 
· 1.14 won five times
· 1.95z lost three times

The ten ABC/HR log files (with reliable ABX results on the first tests) can be found here :
http://membres.lycos.fr/guruboolez/AUDIO/samples/Tosca

As we can see, 1.95z67 seems to be the worse encoder for this (dual) sample. I often designed this encoder as the worse (7 times). Usually, 1.95z67 encodings were the first to be unmasked... So, there's no doubt : 1.95z67 has more difficulties than the latest SV7 codec.

More interesting, the 1.14 vs 1.15r fight. During the first two seconds, 1.15r showed its superiority (4 times on five test). But two seconds later, results are inverted : 1.14 won, five times (on five), and 1.15r was designed as the worst two times.

=> On a same problem (same recording, same instrument, same minute), two mpc encoders can't keep the same transparency level.

(P.S. This classical sample could feed any future Public blind listening test, even at high bitrate)

EDIT : upload finished.

(slight) Problem sample

Reply #1 – 2003-08-04 03:36:44

OMG! He did it again. Guruboolez ABXd Musepack -insane.

Hey, Guru, where can I get ears like yours?

(slight) Problem sample

Reply #2 – 2003-08-04 13:32:49

This sample can be ABXd with 1.15r at --insane at around 0.6 seconds. There is indeed some distortion, on the left channel. Somewhere between --quality 7.1 and 7.5 the problem goes away, but it is possible that it can still be heard by some, even at --quality >7.5.

Notice