Hi,

I've found another problem sample for mpc 1.14 --standard (and other lossy codecs). The track is Basscadet from Autechre's first album 'Incunabula'. I immediately suspected a difference, so I did an abx-test. Results: first round 7/8, second round 8/8, third round 8/8. I didn't test higher bitrate-modes because standard is supposed to be transparent at that setting, and I was getting bored with further testing. The problem seems that the distorted 'hihats' sounds last longer in the mpc sample than in the original, and a little more 'coarse' sound, which all seems similar to other mpc-problem samples posted here recently, and I also suspect a problem with the stereo-imaging, I don't know how to describe it.

The sample could be of further interest for lame -aps tuning, because lame -aps fails completely with this sample, far more worse than 2nd Bad Vilbel sample, also by Autechre.

Here is a (short) original sample.

Ae_Basscadet.flac

I think it would be in the interest of all mpc-users here if other people could confirm this problem with further abx-results.

-tm

I tried many time, and I don't ear anything wrong with --standard. Can you tell me the exact passage you ABXed with succes ?

Not exactly, for me it's in the hole sample, after each 'Tschack-tschack-tschack-tshack', it doesn't help for to me single out one portion, I hear it (it is not an obvious artifact) at the end of each 'Tschack-tschack-tschack-tshack', it 'feels' like the last 'Tschack' last longer, like a post-echo, subtle, but abx-able (did it again 5/5)

Maybe it would help to listen to the original a long time, 'flush' memory, and then listen to the encoded sample, like you did with the castanets sample.

Maybe I spotted it because I know this track for about seven years now...

O, and I listen with headphones (Sennheiser HD 540 reference II on SB Live value)

Keep trying,


-tm

I will try in seven years
No, sorry, but the only thing I feel is that I'm going to be crazy if I continue to listen to it.

Ah, I hear it. First ABX 6/16 (wasn't really an ABX, just listening to the sample and guessing)

Then I started hearing it and quickly ABX'd 14/16.

At the beginning of the clip (first 1.9 seconds), there's a pattern:

t-----t-t-t-----t-----T-T-T-----T-----

where I marked the capital T's, the --standard preset seems to "lower the pitch" of the attacks and, as Tim Mervielde said, soften the end of the attacks.

I did a quick ABX of --xtreme and failed, but I think I still hear the problem... I will continue testing to see if I'm just imagining things.

(using mppenc v1.14, default settings, and mppdec v1.93j, default settings)

I listened to it again, and I find the second part to be a bit different.
I ABXed it again, focused on 5.1-5.7. Scores are sometimes very good (around 8-9/10) and sometime insignificant. Maybe luck, maybe concentration lost.
Must try tomorrow...

But it seems that a small & audible difference exists.


P.S. This is not the first time that ABX score are playing yo-yo by testing mpc --standard. I did some tests with samples I just suspect to be difficult to encode. Sometimes ABX are good, and just after, they are calamitous. Doesn't know if only luck can explain that.

guruboolez, for a more reliable result and conclusion over multiple ABX tests, add all the trials and correct responses, and calculate the resulting p-value. You can calculate it at ff123's p-value calculator web page. There's also an excel table at my web page.

I did some further testing...

Results:
*When I downmix the sample to mono, I can't Abx it any more (default settings)

*When I use the --ms 15 switch with the standard settings, I can abx it, but it's more difficult:

02:24:47 0/1 p=100.0%
02:25:05 1/2 p=75.0%
02:25:46 2/3 p=50.0%
02:26:16 3/4 p=31.2%
02:26:32 4/5 p=18.8%
02:27:21 5/6 p=10.9%
02:27:40 5/7 p=22.7%
02:28:13 5/8 p=36.3%
02:28:25 5/9 p=50.0%
02:29:19 6/10 p=37.7%
02:29:33 7/11 p=27.4%
02:30:31 7/12 p=38.7%
02:30:41 8/13 p=29.1%
02:31:18 9/14 p=21.2%
02:31:51 9/15 p=30.4%
02:32:39 10/16 p=22.7%
02:32:57 11/17 p=16.6%
02:33:21 12/18 p=11.9%
02:33:39 13/19 p= 8.4%
02:35:01 14/20 p= 5.8%
02:35:27 14/21 p= 9.5%
02:35:53 15/22 p= 6.7%
02:36:08 16/23 p= 4.7%
02:36:27 17/24 p= 3.2%
02:36:38 test finished

* I cannot abx it with the -ms 0 switch in standard (I know, ms 0 is supposed to be not-done)

* But I can still abx it with the default standard -ms 11 settings, which means it's not a problem of concentration

So just maybe there's a problem when using mid/side coding, but I can't abx standard -ms 11 against standard -ms 0


-tm

I tried xtreme. it wasn't easy but 4/4 ABA p=1.2%

I would still rate the quality at xtreme as very good.. Didn't test standard quality, but I'm not sure if these kind of extreme clips can ever be transparent at standard level..

Are you suggesting that: 1) the tuning needed to make Musepack recognize and fix non-transparent samples like this one would be very difficult, or 2) correcting for this sample would bloat bitrates for all music?

Or that standard's real-world objective is, say, 99.9% transparency, as opposed to 99.95%? (this is of course different from the "goal" of 100% transparency)

Or none of the above?