Here's a new killer sample for MPC : the worst I know, worse than Amnesia :

http://perso.numericable.fr/laguill2/files/fsol.flac

It's Future Sound Of London, Lifeforms, Among Myselves. I noticed something wrong listening to it, and confirmed comparing to the original CD. And it doesn't clip.

MPPENC (1.14) --xlevel --quality 5

ABX 50/50, soooo easy !

Quality 7 : ABX 8/8
Quality 8 : ABX 8/8
Quality 9 : ABX 6/8 it begins to be solved.

Lame 3.90.3 --alt-preset standard is not annoying. At least as good as MPC quality 8.

A strange listening fatigue showed up with this sample. I first noticed immediately the problem at quality 8. Then, I became confused... I couldn't even distinguish quality 7 from the original. So I rested my ears.
Then I could ABX quality 7 without problem.
Then, quality 8 was more difficult. I had to separate the sessions with 10 seconds of silence and listen only to X to avoid listening fatigue, showing up after two or three listenings only !

What is very surprising is the obviousness of the Q8 problems when it is directly played after 1 minute of silence, while after having listening to the sample 30 times in a row, Q7 doesn't even sound different from the original anymore.

It is an interesting observation about ABX tests for me. When I detect a problem casually listening to my files, I should be more careful at ABX ing it after. I'm now sure that I missed some ABX tests that I could have succeeded because of this listening fatigue.

Note : the results given above are all the results I got, and the number of trials was fixed before. When I say that I couln't distinguish the files, I didn't answer the ABX question. I only gave a choice (A or B) when I was sure of it.

I was just curious and did a quick ABX for myself.
But it seems I'm a not in a excellent form for listing tests.

I scored with

MPC 1.14 -Q5 11/14

Lame 3.90.3 -APS 12/16

Will try again tomorrow.

Can you please tell where exactly the difference was most obvious to you?

0.3 - 1.0 sounds quite bad to me with 1.14 q5 and 1.15r q5 (both 8/8), but at q7 and higher I'd have to spend some time to get more familiar with the sample to have a chance.

0.2 - 1.1: ABXed successfully (p-value < 0.01) 1.14 q5, q7, q8 and 1.15r q5, q7 (haven't tried q8 or higher). At q5 1.14 and 1.15r seem to sound different, but I can't tell what sounds closer to original.

For me q5 is really obvious + awful, ABXing without fatigue possible.
At q7 the difference I focus on first gets lost after a few trials, but it's easy to find another one. With q8 I need a break after a few trials. With q7 or 8 I don't think I'd ever notice anything wrong outside of ABX situation, but probably because I'm not familiar with this kind of music.

Seems it wasn't a good idea from me to ABX the whole sample.

Focusing on part 0.3 - 1.1 sec is really interesting.

It was easy to ABX at MPC 1.14 -Q5 with result 12/12

But there is also a ABX'able difference between 1.5 and 2.5 sec for my ears but not that much

From 0.4 s to 1.5 s.

EDIT : The most obvious artifact for me is between 1.0 and 1.3 s.

ABX'd Q7 8/8 without much effort, MPC 1.14.

Not yet tried Q8 or mp3 aps.

Very interesting sample that Pio, thanks!

Edit:


Yeah, don't think I would either.

At --standard, the first artifact I noticed was the later one one noticed by naturfreak. I had pretty bad problems becoming fatigued with it though. I would get the first 6 or 7 trials right, then would start suddenly missing quite a few and couldn't hear it at all beyond 15 or 16 trials. If I rested a few minutes, I could easily hear it again. Still I was able to get 12 to 14 out of 16 each time. The earlier part which I didn't notice clearly until other's mentioned where it was was much easier. Got 9/10 and then 10/10. Got 9/10 at -q6. Haven't gotten it at -q7 yet, but my computer is very noisy.

I would describe the artifacts as noticeable but not annoying at --standard. I thought waiting.wav was a bit easier at --standard, but I couldn't ABX that one at -q6 (though I'm more trained now and might be able to if I try again.)

PIO could you test 1.15r with q7? be sure to post if it improves on 1.14b

I found it harder to detect artifacts with 1.14b at q5.

Too bad I didn't see this thread back then.

This sample has very dense (28 ms inbetween) pulses that cause MPC and Nero's
AAC encoder to fail. 1.14 and 1.15r are similarly affected and easy to ABX.

WinABX v0.42 test report
02/24/2004 20:25:02

mppenc 1.15r --quality 5:

20:26:32 1/1 p=50.0%
20:27:43 2/2 p=25.0%
20:33:05 3/3 p=12.5%
20:33:27 4/4 p=6.2%
20:34:24 5/5 p=3.1%
20:34:59 6/6 p=1.6%

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

1.15r, --quality 6:

WinABX v0.42 test report
02/24/2004 20:36:09

A file: E:\fsol.wav
B file: E:\fsol-mpcq6.wav

Start position 00:00.4, end position 00:01.5
20:38:21 1/1 p=50.0%
20:38:39 2/2 p=25.0%
20:38:54 3/3 p=12.5%
20:39:11 4/4 p=6.2%

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

Nero AAC latest - Extreme

WinABX v0.42 test report
02/24/2004 21:08:52

A file: E:\fsol.wav
B file: E:\fsol-aac-extreme.wav

Start position 00:00.4, end position 00:01.5
21:10:31 1/1 p=50.0%
21:11:26 2/2 p=25.0%
21:13:12 3/3 p=12.5%
21:15:50 4/4 p=6.2%

This sample hurts my ears. There's no doubt tuning is needed on both codecs to handle this. I don't have time to try at higher qualities. The sample should be added to the collection, along with the other samples we have. In any case, you guys should provide logs, rather than just say "I did it at X/X". This does not help others compare their attempts to your attempts.

Why? Do you want to compare how good other people started, are you interested in the used time?

I definitely prefer a description to extensive logs. I'm particulary not interested in Pio's 50/50 test...

It would be helpful to know the time between attempts and how long each tester needs to rest. I found I needed a lot of rest between attempts with 1.15r. With 1.14 and AAC I preferred to do it quickly with very little rest. I will try MPC --insane and Nero's "Audiophile" later, when it's quieter.

Easy ABXable, 10/10.

mppenc 1.15r --quality 5 --xlevel

I'll try now --quality 6, 7, 8 and try this as well with mppenc 1.14

WinABX v0.42 test report
02/25/2004 00:05:50

Difficult, but there is still a bit of smearing and pre/post echo
that causes the transients to lose their crunchiness.

A file: E:\fsol.wav
B file: E:\fsol-AAC-Audiophile.wav

00:11:57 1/1 p=50.0%
00:12:15 2/2 p=25.0%
00:12:56 3/3 p=12.5%
00:18:37 3/4 p=31.2%
00:19:24 4/5 p=18.8%
00:22:52 5/6 p=10.9%
00:25:15 6/7 p=6.2%

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

With 1.15r --insane, I was only able to hear a difference when I lowered the volume to ~40%. It caused the higher frequencies to almost disappear, and
then it was slighty easier to hear the smearing at the lower frequencies. At
a higher volume I failed to ABX this sample over and over again. Seems
like masking plays a big role in this sample. Do you all know more or less
how high you can hear? It's interesting.

mppenc 1.15r --insane

WinABX v0.42 test report
02/25/2004 00:28:22

A file: E:\fsol.wav
B file: E:\fsol-q7.wav

Start position 00:00.4, end position 00:01.5
00:32:22 0/1 p=100.0%
00:32:25 reset

00:33:44 1/1 p=50.0%
00:37:24 1/2 p=75.0%
00:37:53 1/3 p=87.5%
00:37:57 reset

00:39:48 0/1 p=100.0%
00:39:49 reset

00:41:24 1/1 p=50.0%
00:41:36 2/2 p=25.0%
00:42:25 2/3 p=50.0%
00:42:54 3/4 p=31.2%
00:43:38 4/5 p=18.8%
00:45:00 5/6 p=10.9%
00:45:51 6/7 p=6.2%
00:46:23 7/8 p=3.5%
00:47:13 7/9 p=9.0%

I lowered the volume a little bit more and tried to concentrate as much as I could:

Nero AAC encoder (latest version), "Transcoding":

WinABX v0.42 test report
02/25/2004 01:07:53

A file: E:\fsol.wav
B file: E:\fsol-AAC-Transcoding.wav

01:16:14 1/1 p=50.0%
01:16:26 2/2 p=25.0%
01:16:58 3/3 p=12.5%
01:17:25 4/4 p=6.2%
01:19:49 reset

01:23:00 1/1 p=50.0%
01:23:44 2/2 p=25.0%
01:24:54 3/3 p=12.5%
01:26:05 4/4 p=6.2%
01:28:43 reset

Pre-echo still seems to exist, but it's extremely hard to detect now.

I find this sample easier to ABX than 'Waiting'.

mppenc --xlevel --standard

The sample consist in a frequency going up and down very fast, rather than pulses. Here is a detailed somogram. The picture is about 300 ms wide.

The sonogram of the MPC 1.14 standard version exhibits drop outs as low as 2 kHz !

About my tests, I didn't post logs because I don't have any. PCABX don't generate logs. But the 50/50 was done very, very fast. The speed limiting factor was the mouse clicking on the right button. I recognized X instantly (as soon as the dropped out part was played) without comparing to A or B.