Lame 3.98 - transients test

Topic: Lame 3.98 - transients test (Read 8171 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.

Lame 3.98 - transients test

2005-12-04 16:53:00

http://gabriel.mp3-tech.org/lame/lame_attack.exe

I would be pleased to receive feedback regarding how this version behaves regarding transients under the following two conditions:
*target bitrate at least 112kbps or at least V5
*sampling freq 44.1kHz

Thanks

Lame 3.98 - transients test

Reply #1 – 2005-12-04 19:23:10

As far as I'm concerned, I can't hear a difference; Somehow, I just found out that my Lame 3.97beta is a 3.98 alpha, though.. odd..

Here's a first abx test result :

(samples here)

Code: [Select]

lame_attack -V 4 --vbr-new "Aerial.wav" "Aerial.wav"-attack4.mp3

LAME 3.98 (alpha 2, Dec  4 2005 17:29:32) 32bits (http://www.mp3dev.org/)
warning: alpha versions should be used for testing only
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used)
Using polyphase lowpass filter, transition band: 17249 Hz - 17782 Hz
Encoding as 44.1 kHz VBR(q=4) j-stereo MPEG-1 Layer III (ca. 10x) qval=3
    Frame          |  CPU time/estim | REAL time/estim | play/CPU |    ETA
 10890/10890 (100%)|    0:22/    0:22|    0:22/    0:22|   12.513x|    0:00
 32 [  172] %***
 40 [    1] *
 48 [    2] *
 56 [    3] *
 64 [    9] *
 80 [   38] *
 96 [  295] ******
112 [ 1844] %********************************
128 [ 3109] %*******************************************************
160 [ 3769] %******************************************************************
192 [  771] %*************
224 [  561] %*********
256 [  199] %***
320 [  117] %**
-------------------------------------------------------------------------------
   kbps        LR    MS  %     long switch short %
  147.6        0.7  99.3        98.4   0.9   0.7

lame_attack -V 5 --vbr-new Aerial.wav Aerial.wav-attack5.mp3

LAME 3.98 (alpha 2, Dec  4 2005 17:29:32) 32bits (http://www.mp3dev.org/)
warning: alpha versions should be used for testing only
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used)
Using polyphase lowpass filter, transition band: 15826 Hz - 16360 Hz
Encoding as 44.1 kHz VBR(q=5) j-stereo MPEG-1 Layer III (ca. 11.9x) qval=3
    Frame          |  CPU time/estim | REAL time/estim | play/CPU |    ETA
 10890/10890 (100%)|    0:22/    0:22|    0:22/    0:22|   12.758x|    0:00
 32 [  190] %**
 40 [    3] *
 48 [    5] *
 56 [   16] *
 64 [   16] *
 80 [  168] ***
 96 [ 1866] ****************************
112 [ 4513] %******************************************************************
128 [ 2012] %*****************************
160 [ 1049] %***************
192 [  520] %*******
224 [  349] %*****
256 [  111] %*
320 [   72] %*
-------------------------------------------------------------------------------
   kbps        LR    MS  %     long switch short %
  125.0        0.7  99.3        98.5   0.8   0.7

lame -V 4 --vbr-new Aerial.wav Aerial.wav-4.mp3

LAME 3.98 (alpha 2, Nov 23 2005 14:57:40) 32bits (http://www.mp3dev.org/)
warning: alpha versions should be used for testing only
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used)
Using polyphase lowpass filter, transition band: 17249 Hz - 17782 Hz
Encoding as 44.1 kHz VBR(q=4) j-stereo MPEG-1 Layer III (ca. 10x) qval=3
    Frame          |  CPU time/estim | REAL time/estim | play/CPU |    ETA
 10890/10890 (100%)|    0:36/    0:36|    0:36/    0:36|   7.8575x|    0:00
 32 [  172] %**
 40 [    0]
 48 [    3] *
 56 [    1] *
 64 [    1] *
 80 [    9] *
 96 [  287] *****
112 [ 1934] %*********************************
128 [ 3269] %********************************************************
160 [ 3873] %******************************************************************
192 [  745] %************
224 [  476] %********
256 [   96] **
320 [   24] *
-------------------------------------------------------------------------------
   kbps        LR    MS  %     long switch short %
  144.2        0.7  99.3        88.2   6.7   5.1

lame -V 5 --vbr-new Aerial.wav Aerial.wav-5.mp3

LAME 3.98 (alpha 2, Nov 23 2005 14:57:40) 32bits (http://www.mp3dev.org/)
warning: alpha versions should be used for testing only
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used)
Using polyphase lowpass filter, transition band: 15826 Hz - 16360 Hz
Encoding as 44.1 kHz VBR(q=5) j-stereo MPEG-1 Layer III (ca. 11.9x) qval=3
    Frame          |  CPU time/estim | REAL time/estim | play/CPU |    ETA
 10890/10890 (100%)|    0:41/    0:41|    0:41/    0:41|   6.8444x|    0:00
 32 [  189] %**
 40 [    4] *
 48 [    2] *
 56 [    2] *
 64 [    4] *
 80 [  137] **
 96 [ 2102] ******************************
112 [ 4786] %******************************************************************
128 [ 1988] %***************************
160 [ 1141] %***************
192 [  420] %*****
224 [   95] **
256 [   16] *
320 [    4] *
-------------------------------------------------------------------------------
   kbps        LR    MS  %     long switch short %
  119.4        0.6  99.4        89.1   6.2   4.7

Code: [Select]

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/04 13:52:30

File A: 01. Michael Hedges - Aerial Boundaries.wv
File B: Aerial.wav-attack4.mp3

13:52:30 : Test started.
13:53:47 : 01/01  50.0%
13:53:55 : 02/02  25.0%
13:54:04 : 03/03  12.5%
13:54:13 : 04/04  6.3%
13:54:25 : 05/05  3.1%
13:55:09 : 05/06  10.9%
13:55:12 : 05/07  22.7%
13:55:16 : 06/08  14.5%
13:55:26 : 06/09  25.4%
13:55:27 : Test finished.

 ---------- 
Total: 6/9 (25.4%)

Code: [Select]

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/04 13:55:58

File A: 01. Michael Hedges - Aerial Boundaries.wv
File B: Aerial.wav-attack5.mp3

13:55:58 : Test started.
13:57:35 : 00/01  100.0%
13:57:38 : 01/02  75.0%
13:57:44 : 02/03  50.0%
13:57:49 : 02/04  68.8%
13:58:02 : 02/05  81.3%
13:58:10 : 03/06  65.6%
13:58:16 : 03/07  77.3%
13:58:21 : 04/08  63.7%
13:58:23 : Test finished.

 ---------- 
Total: 4/8 (63.7%)

Code: [Select]

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/04 13:58:45

File A: 01. Michael Hedges - Aerial Boundaries.wv
File B: Aerial.wav-5.mp3

13:58:45 : Test started.
14:00:22 : 01/01  50.0%
14:00:28 : 02/02  25.0%
14:00:37 : 03/03  12.5%
14:00:41 : 03/04  31.3%
14:00:44 : 03/05  50.0%
14:01:00 : 03/06  65.6%
14:01:03 : 04/07  50.0%
14:01:24 : 05/08  36.3%
14:02:11 : 06/09  25.4%
14:02:11 : Test finished.

 ---------- 
Total: 6/9 (25.4%)

Code: [Select]

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/04 14:02:37

File A: 01. Michael Hedges - Aerial Boundaries.wv
File B: Aerial.wav-4.mp3

14:02:37 : Test started.
14:04:05 : 01/01  50.0%
14:04:22 : 02/02  25.0%
14:04:28 : 02/03  50.0%
14:04:36 : 02/04  68.8%
14:04:38 : 02/05  81.3%
14:04:41 : 03/06  65.6%
14:04:48 : 03/07  77.3%
14:04:57 : 04/08  63.7%
14:05:12 : 04/09  74.6%
14:05:16 : 04/10  82.8%
14:05:20 : 04/11  88.7%
14:05:26 : 05/12  80.6%
14:05:33 : 05/13  86.7%
14:05:33 : Test finished.

 ---------- 
Total: 5/13 (86.7%)

Lame 3.98 - transients test

Reply #2 – 2005-12-04 19:38:48

I concurr, they're all indiscernable, to me. Anyone can find the samples here to ABX. Mp3s included.

Lame 3.98 - transients test

Reply #3 – 2005-12-04 20:17:18

I have no transient results but i want to add, that this lame_attack.exe behaves better with the velvet sample. Using -V2 --vbr-new the problems descriped here are not that obvious anymore.
http://www.hydrogenaudio.org/forums/index....ndpost&p=346967
Now the pumping is only added in the very beginning and from second 7-9. The very loud pumping at seconds 9-10 is gone.

Lame 3.98 - transients test

Reply #4 – 2005-12-04 21:03:03

If the good old awe32-20sec sample is a so called transient sample i have one result.
lame_attack.exe -V2 --vbr-new adds noise like silent hi-hats from seconds 0-6. 397b doesn´t.
Easy to hear.
If this old sample isn´t available anymore and someone likes it i may upload it.

Lame 3.98 - transients test

Reply #5 – 2005-12-04 21:15:33

I think the trumpet sample is a transient sample if I remember correctly Xin (the one with the headphone amps) writing some day that female voice, piano and trumpet are difficult sources for electronic equipment - the trumpet because of transients.

Good news: I could not abx the trumpet sample with lame_attack -b320 -h, and I don't think this is because my ears may be a bit tired because I've just returned from a loud concert (but I'll repeat the test tomorrow).

Moreover I've noticed lame_attack uses the full 511 Byte of bit reservoir.
Edited:
I was wrong saying full 511 Byte are used for bit reservoir. For cbr320 it's still 396 Byte. 511 Byte are used with cbr256 (I did a short cbr256 test before really abxing cbr320).

Thank you, Gabriel.

Lame 3.98 - transients test

Reply #6 – 2005-12-04 21:19:29

I did a small test using giveuptheghost-sincealways.sample18sec.flac.

I tested -V5 --vbr-new with 3.97b2 and lame_attack:

Code: [Select]

ABC/HR Version 1.0, 6 May 2004
Testname: giveuptheghost-sincealways Listening Test [128kbps]

1L = samples\dec\giveuptheghost-sincealways.sample18sec.lame_attack.wav
2R = samples\dec\giveuptheghost-sincealways.sample18sec.397b2.wav

---------------------------------------
General Comments:

---------------------------------------
1L File: samples\dec\giveuptheghost-sincealways.sample18sec.lame_attack.wav
1L Rating: 4.3
1L Comment: 
---------------------------------------
2R File: samples\dec\giveuptheghost-sincealways.sample18sec.397b2.wav
2R Rating: 4.0
2R Comment: Slightly worse Pre-Echo (Especially notable: 0.0-2.5)
---------------------------------------
ABX Results:
Original vs samples\dec\giveuptheghost-sincealways.sample18sec.lame_attack.wav
    9 out of 9, pval = 0.002
Original vs samples\dec\giveuptheghost-sincealways.sample18sec.397b2.wav
    9 out of 9, pval = 0.002
samples\dec\giveuptheghost-sincealways.sample18sec.lame_attack.wav vs samples\dec\giveuptheghost-sincealways.sample18sec.397b2.wav
    7 out of 9, pval = 0.090

The difference between the two was slight but notable, but I'm very insecure about my preference of lame_attack. If anybody could give it a shot, I'd be greatful.

Lame 3.98 - transients test

Reply #7 – 2005-12-04 21:36:09

Quote

... awe32-20sec sample ...
If this old sample isn´t available anymore and someone likes it i may upload it.
[a href="index.php?act=findpost&pid=347649"][{POST_SNAPBACK}][/a]

I was searching for it without success, so I thank you very much if you will upload it.

Lame 3.98 - transients test

Reply #8 – 2005-12-05 21:31:25

When I tested lame_attack on the trumpet sample yesterday I was a bit uncertain whether my listening abilities might have suffered from the concert I visited before.

I just redid the test and can confirm: I am not able to abx the trumpet sample with lame_attack -b320 -h.

Edited:
But I was wrong saying full 511 Byte are used for bit reservoir. For cbr320 it's still 396 Byte. 511 Byte are used with cbr256 (I did a short cbr256 test before really abxing cbr320).

Looks like the limit on audio content buffer (according to a 32 kHz sampled cbr320 transport frame) is implemented as max audio content of current transport frame + a limit to bit reservoir. This gives a bit reservoir limit of actually 396 Byte for cbr320 resp. 511 Byte for cbr256.
But why not take audio content buffer limit in its own right and always allow the full 511 Byte to bit reservoir?

Lame 3.98 - transients test

Reply #9 – 2005-12-08 08:49:44

I'm bumping this, because I want someone to confirm my results. It's an easy sample, though I wouldn't consider it a problem sample.

Lame 3.98 - transients test

Reply #10 – 2005-12-08 10:41:59

Quote

I'm bumping this, because I want someone to confirm my results. It's an easy sample, though I wouldn't consider it a problem sample.
[a href="index.php?act=findpost&pid=348567"][{POST_SNAPBACK}][/a]

I'm not getting anything from your sample except a headache, dev0

05:38:16 : Test started. (NORMAL)
05:39:08 : 01/01 50.0%
05:39:17 : 02/02 25.0%
05:39:24 : 02/03 50.0%
05:39:27 : 03/04 31.3%
05:39:29 : 03/05 50.0%
05:39:34 : 03/06 65.6%
05:39:42 : 03/07 77.3%
05:39:46 : 04/08 63.7%
05:39:48 : Test finished.

05:36:05 : Test started. (ATTACK)
05:37:02 : 00/01 100.0%
05:37:09 : 01/02 75.0%
05:37:13 : 01/03 87.5%
05:37:19 : 02/04 68.8%
05:37:24 : 03/05 50.0%
05:37:27 : 03/06 65.6%
05:37:42 : 04/07 50.0%
05:37:51 : 04/08 63.7%
05:38:04 : 05/09 50.0%
05:38:05 : Test finished.

One sample seemed to have slightly less loud hat hits than the other, but I can't really say that with.. any degree of confidence at all, can I ?

Lame 3.98 - transients test

Reply #11 – 2005-12-08 14:08:08

I can't either. Did you use -V5 or something higher?

Lame 3.98 - transients test

Reply #12 – 2005-12-08 14:39:47

Quote

I can't either. Did you use -V5 or something higher?
[a href="index.php?act=findpost&pid=348614"][{POST_SNAPBACK}][/a]

Same settings as you, to try to confirm : -V 5 --vbr-new. I'm ashamed to have good audio equipment, but to be unable to benefit from it...

BTW : I really like your avatar. Did you draw it yourself?

Lame 3.98 - transients test

Reply #13 – 2005-12-11 20:57:49

I tried lame_attack on the samples from the problem sample thread.
I used
- Birds
- Moon
- Deploration
- herding calls
- trumpet
and could easily hear the distortions with 3.97b2 -V2 --vbr-new.
I also wanted to use aps_killer and velvet but was not able to hear the problems. So I left them over.

lame_attack -V2 improved remarkable on the samples but only when not using --vbr-new.

With lame_attack -V2 to me only herding calls and trumpet remained as showing up problems.
No VBR mode could solve the problems with these two samples however, not even -V0 -b256.

lame_attack --preset 256 didn't solve the problems either, but with additional restrictions on bit rate and/or increasing average bit rate results became acceptable.
With lame_attack --preset 256 -b224 or --preset 270 -b256 the trumpet sample became transparent to me, and herding calls was hard to abx.
Herding calls did not become transparent even with lame_attack -b320 -h (10/10), though it was extremely hard to abx.

Taking it all together for the examples and circumstances considered
- lame_attack yields a remarkable quality increase compared to 3.97b2
- --vbr-new yields worse results than --vbr-old
- quality control is insufficient with VBR
- brute force ABR or CBR yield quite satisfying though not perfect results at bit rates around 256 kbps.
- ABR should be used with restrictions on minimum bit rate.
- lame_attack is not transparent on herding calls with cbr320.

For a comparison whether it's only the samples or also the encoding machinery that bring the problems I tried Lame 3.88b on the samples using -b224 -h. Though I felt not all of the encodings are transparent I could not abx one of them against the original. (Maybe I'm used to the 3.97 distortions which are not present in the 3.88 encodings and may be I am not sensitive at the moment for other kind of artefacts).

Lame 3.98 - transients test

Reply #14 – 2005-12-15 09:23:57

I quickly tried lame_attack on several (40) pre-echo samples with --V5 --vbr-new. I compared it with 3.97b2 with the same setting:
• 3.97 beta 2 --V5 --vbr-new
• 3.98_attack --V5 --vbr-new

When I say quickly, it’s really quickly. Quick comparisons are not suitable for finding subtle difference. That’s maybe why I didn’t notice any improvements. Smeared samples with 3.97 are still smeared with 3.98_alpha. Maybe less, maybe more: if a difference exists, I didn’t noticed it. But I noticed four samples which offered obvious difference. All correspond to a similar and known short-impulse/micro-attacks situation. Unfortunately, the difference is in favor of 3.97b2: 3.98 introduced new artifacts and lowered the quality.

• atem-lied: warbling (« plop-bloup-bloup » inside). Audible during the 00.00 – 06.00 range and with 3.98_attack. 3.97 beta has the same issue, but it’s hardly noticeable in comparison. Side note: there are several big vertical holes in the frequency representation and these missing information correspond to the moment when artifacts occur.

• awe32_20: sound is noisier on several parts: 01.200 – 01.800 ; 04.100 – 05.200 ; 06.200 – 07.200 ; 07.900 – 11.500. It’s like a change in tonality. 3.97b2 has no similar problem, but has audible smearing (same for 3.98_attack)
Side note: there is a big disparity in bitrate allocation – the biggest for all four samples. 3.97 used 180 kbps whereas 3.98 only requested 153 kbps (encspot values). 80% of short block with the beta version; “only” 57% for the experimental one.

• cataclysmes: both encoders are unable to reproduce the precise details: edges are totally blurred. 3.98_attack has more pre-echo/noise.

• creaking: like awe32_20 there are small parts which are noisier and which sound like a different tonality. Parts are: 0.900 – 1.100 and 1.200 – 1.400. During the 0.900 – 1.400 range, two artifacts occur, easy to hear in my opinion.

I created four archives, including reference file (lossless) and both encodings. It would be nice to see additional comments for these files.

EDIT: archives are temporary available here.

Lame 3.98 - transients test

Reply #15 – 2005-12-15 20:09:43

Thanks

Lame 3.98 - transients test

Reply #16 – 2005-12-16 20:23:51

I tried Gurus samples.
I have great difficulties hearing problems with electronic music or non-instrument resp. non-vocal samples.
I can confirm hearing the Lame_attack -V5 artefacts on creaking, and to me too they are a worse than those produced by 3.97b2. With Lame_attack -V2 however I could not reliably abx them..

Best for me to hear the problems was Atem-lied.
I could abx Lame_attack -V2.

I don't have a lot of confidence in VBR since quite a while when it comes to difficult samples. So I tried Lame_atack --preset 240 -b192 -h as well as Lame 3.88b and 3.91 (--abr 240 -b192 -h). All three encodings were not transparent to me, with Lame_attack being rather reasy to abx, and 3.88b rather hard.

Next I tried --preset 270 -b224 -h.
Lame_attack problems were again rather easy to abx. 3.88b and 3.91 were transparent to me with --abr 270 -b224 -h.

Notice