Hydrogenaudio Forums > LAME 3.98.2 regression issues at V3

Full Version: LAME 3.98.2 regression issues at V3 - V2.

Hydrogenaudio Forums > Lossy Audio Compression > MP3 > MP3 - Tech

/mnt

Jul 25 2009, 13:11

So far i've been using LAME 3.98.2 at -V2 for over 2 weeks now. But after a while i found some tracks from my lossy collection that are easy to ABX at V2 and sometimes easy to spot artifacts with my iPod earbuds

. Sadly most of the artifacts i've spotted is regression from 3.97.

An example:

LAME 3.98.2 -V2

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/07/25 12:14:44

File A: C:\Music\Albums\ohGr - SunnyPsyOp\02. maJiK.mp3
File B: C:\Rips\Music\ohGr - SunnyPsyOp\02. maJiK.flac

14:44 : Test started.
15:01 : 01/01  50.0%
15:06 : 02/02  25.0%
15:10 : 03/03  12.5%
15:13 : 04/04  6.3%
15:19 : 05/05  3.1%
15:24 : 06/06  1.6%
15:30 : 07/07  0.8%
15:36 : 08/08  0.4%
15:44 : 09/09  0.2%
15:54 : 10/10  0.1%
16:00 : 11/11  0.0%
16:08 : 11/12  0.3%
16:14 : 12/13  0.2%
16:18 : 13/14  0.1%
16:26 : 14/15  0.0%
16:29 : Test finished.

 ---------- 
Total: 14/15 (0.0%)

Nasty distortion at 1:00.

Any way i made a edit version and will soon upload it for anyone else wanting to try the sample out. For some reason the artifact moved on the edit.

LAME 3.98.2 V2 (Edit)

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/07/25 12:32:49

File A: C:\Temp\Samples\maJiK (Edit).flac
File B: C:\Temp\Samples\maJiK (Edit) -V 2 LAME 3.98.2.mp3

32:49 : Test started.
33:04 : 01/01  50.0%
33:13 : 02/02  25.0%
33:16 : 03/03  12.5%
33:20 : 04/04  6.3%
33:24 : 05/05  3.1%
33:29 : 06/06  1.6%
33:34 : 07/07  0.8%
33:38 : 08/08  0.4%
33:43 : 09/09  0.2%
33:48 : 10/10  0.1%
33:53 : 10/11  0.6%
33:57 : 11/12  0.3%
34:02 : 12/13  0.2%
34:10 : 13/14  0.1%
34:17 : 14/15  0.0%
34:26 : 15/16  0.0%
34:27 : Test finished.

 ---------- 
Total: 15/16 (0.0%)

Distortion at 0:07.

LAME 3.97 -V 2 --vbr-new

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/07/25 12:34:41

File A: C:\Temp\Samples\maJiK (Edit).flac
File B: C:\Temp\Samples\maJiK (Edit) -V 2 --vbr-new LAME 3.97.mp3

34:41 : Test started.
35:03 : 01/01  50.0%
35:14 : 01/02  75.0%
35:24 : 02/03  50.0%
35:29 : 02/04  68.8%
35:33 : 02/05  81.3%
35:39 : 03/06  65.6%
35:44 : 04/07  50.0%
35:49 : 05/08  36.3%
35:54 : 05/09  50.0%
36:00 : 06/10  37.7%
36:03 : 07/11  27.4%
36:09 : 08/12  19.4%
36:15 : 09/13  13.3%
36:27 : 10/14  9.0%
36:32 : 11/15  5.9%
36:38 : 12/16  3.8%
36:46 : 13/17  2.5%
36:53 : 14/18  1.5%
37:01 : 14/19  3.2%
37:14 : 15/20  2.1%
37:15 : Test finished.

 ---------- 
Total: 15/20 (2.1%)

Sounds better then 3.98.2.

LAME 3.98.2 vs LAME 3.97 (Sample)

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/07/25 12:38:08

File A: C:\Temp\Samples\maJiK (Edit) -V 2 --vbr-new LAME 3.97.mp3
File B: C:\Temp\Samples\maJiK (Edit) -V 2 LAME 3.98.2.mp3

38:08 : Test started.
38:18 : 01/01  50.0%
38:22 : 02/02  25.0%
38:26 : 03/03  12.5%
38:30 : 04/04  6.3%
38:34 : 05/05  3.1%
38:39 : 06/06  1.6%
38:43 : 07/07  0.8%
38:49 : 08/08  0.4%
38:53 : 09/09  0.2%
38:59 : 10/10  0.1%
39:03 : 11/11  0.0%
39:06 : 12/12  0.0%
39:12 : 13/13  0.0%
39:18 : 14/14  0.0%
39:23 : 15/15  0.0%
39:29 : Test finished.

 ---------- 
Total: 15/15 (0.0%)

LAME 3.97 version sounds better. File B had noticable distortion at 0:07.

I will post some more ABX logs and samples possibly this weekend or early next, depending how much time i have.

EDIT:

MaJiK sample

chichazor

Aug 2 2009, 21:07

Do you test it with 3.99a1? Maybe this regression is gone... I hope...

Rio

Aug 3 2009, 03:47

How about with the ancient 3.90.3?

shadowking

Aug 3 2009, 06:00

QUOTE (Rio @ Aug 3 2009, 12:47)

How about with the ancient 3.90.3?

It usually doesn't work that way . Problems with VBR are typical to all versions - sometimes better or worse depending. In my experience V1 brings some improvements where v3 / v2 have trouble. A high ABR setting or b320 can also provide headroom. Given that space is no longer at such a premium it might make some sense to use V1 or even --abr 287 -h for extra security.

@/mnt: I am interested if a setting like --abr 287 -h will soften or even eliminate these issues ?

memomai

Aug 3 2009, 08:22

anyone else with great ears here for testing this sample and confirm the results of mnt?

kawaiigardiner

Aug 3 2009, 08:42

Could you give me a heads up on the settings you used? I'm using 3.98.2 via XLD with the following settings: -V 0 -q 2

I haven't noticed any artefacts yet - in fact, I've found it as good if not better than Apples own AAC encoder.

halb27

Aug 3 2009, 14:13

No great ears here, but I tested the sample using 3.98.2 -V2 and couldn't hear the problem.

@/mnt:
Is there a problem with -V0?
I'm also interested in the result of shadowking's ABR proposal.

soulsearchingsun

Aug 3 2009, 15:13

This question may be odd (and ot), but could someone please give me a hint where to download 3.97, preferrably the bundled version? -- Thank you.

lvqcl

Aug 3 2009, 16:47

http://homepage.ntlworld.com/jfe1205/LAME/lame3.97.zip

/mnt

Aug 3 2009, 19:57

Some more examples:

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/08/03 09:35:43

File A: E:\Music\Albums\Rage Against The Machine - Evil Empire\02. Bulls On Parade.flac
File B: C:\Temp\Bulls On Parade (V2 LAME 3.98.2).mp3

35:43 : Test started.
36:04 : 01/01  50.0%
36:10 : 02/02  25.0%
36:18 : 03/03  12.5%
36:32 : 03/04  31.3%
36:36 : 04/05  18.8%
36:43 : 05/06  10.9%
36:47 : 06/07  6.3%
36:55 : 07/08  3.5%
37:01 : 08/09  2.0%
37:12 : 08/10  5.5%
37:17 : 09/11  3.3%
37:35 : 10/12  1.9%
37:40 : 11/13  1.1%
37:44 : 12/14  0.6%
37:48 : 13/15  0.4%
37:52 : 14/16  0.2%
37:56 : 15/17  0.1%
38:02 : 16/18  0.1%
38:09 : 17/19  0.0%
38:13 : 18/20  0.0%
38:17 : Test finished.

 ---------- 
Total: 18/20 (0.0%)

Precho at 0:37 - 0:38. Transparent on LAME 3.97 at V2 --vbr-new.

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/08/03 09:39:47

File A: E:\Music\Albums\Skinny Puppy - Ain't It Dead Yet\08. One Time, One Place.flac
File B: C:\Temp\One Time, One Place (V2 LAME 3.98.2).mp3

09:39:47 : Test started.
09:40:00 : 01/01  50.0%
09:40:05 : 02/02  25.0%
09:40:10 : 03/03  12.5%
09:40:14 : 04/04  6.3%
09:40:19 : 05/05  3.1%
09:40:23 : 06/06  1.6%
09:40:30 : 07/07  0.8%
09:40:42 : 08/08  0.4%
09:40:47 : 09/09  0.2%
09:40:58 : 10/10  0.1%
09:41:05 : 11/11  0.0%
09:41:20 : 12/12  0.0%
09:41:22 : Test finished.

 ---------- 
Total: 12/12 (0.0%)

Flanging warbling type artifact at 1:15. Sounds fine on LAME 3.97 at -V2 --vbr-new.

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/08/03 19:21:10

File A: E:\Music\Albums\ohGr - SunnyPsyOp\01. HiLO.flac
File B: C:\Temp\HiLO (LAME 3.98.2 V2).mp3

19:21:10 : Test started.
19:21:30 : 01/01  50.0%
19:21:34 : 02/02  25.0%
19:21:45 : 03/03  12.5%
19:21:51 : 04/04  6.3%
19:21:59 : 05/05  3.1%
19:22:04 : 06/06  1.6%
19:22:11 : 07/07  0.8%
19:22:19 : 08/08  0.4%
19:22:24 : 09/09  0.2%
19:22:28 : 10/10  0.1%
19:22:36 : Test finished.

 ---------- 
Total: 10/10 (0.1%)

Wet precho at 1:18. This track is ABXable with 3.97 (at the start) though.

CODE

foo_abx 1.3.4 report
foobar2000 v0.9.6.8
2009/08/03 19:29:55

File A: C:\Temp\HiLO (LAME 3.98.2 V2).mp3
File B: C:\Temp\HiLO (LAME 3.97 -V2 --vbr-new).mp3

19:29:55 : Test started.
19:30:30 : 01/01  50.0%
19:30:37 : 02/02  25.0%
19:30:42 : 03/03  12.5%
19:30:46 : 04/04  6.3%
19:30:51 : 05/05  3.1%
19:30:59 : 06/06  1.6%
19:31:06 : 07/07  0.8%
19:31:11 : 08/08  0.4%
19:31:15 : 09/09  0.2%
19:31:26 : 10/10  0.1%
19:31:31 : 11/11  0.0%
19:31:36 : 12/12  0.0%
19:31:37 : Test finished.

 ---------- 
Total: 12/12 (0.0%)

Both versions sound abit different at the start, but the annoying precho at 1:18 is not present on the 3.97 encode.

QUOTE (shadowking @ Aug 3 2009, 06:00)

In my experience V1 brings some improvements where v3 / v2 have trouble. A high ABR setting or b320 can also provide headroom. Given that space is no longer at such a premium it might make some sense to use V1 or even --abr 287 -h for extra security.

@/mnt: I am interested if a setting like --abr 287 -h will soften or even eliminate these issues ?

I agree that V1 does make improvements.

I could try them at ABR, but those issues do go away at -V1 and -V0 most of the time.

QUOTE (halb27 @ Aug 3 2009, 14:13)

@/mnt:
Is there a problem with -V0?
I'm also interested in the result of shadowking's ABR proposal.

I will give the ABR setting a go.

Most of my regression issues do go away at V0, but the bitrates that LAME 3.98.2 produces at V0 is HUGE (256 - 300 most of time).

halb27

Aug 3 2009, 20:35

Thanks for testing at -V1 and -V0, /mnt.
I'm glad to hear that the most of the problems are gone with -V0.
Extremely high bitrate can be tamed by using -Y, or --lowpass 16.0 or 16.7 or 17.5 according to personal preference, and/or using mp3packer on the Lame results.
I'm looking forward to your very high bitrate ABR results.

Case

Aug 3 2009, 20:38

Just a quick question, are you still using that Sound Blaster Audigy SE quick googling found you mentioned? If so, it would be best if you used 48 kHz resampling when you test.

Squeller

Aug 3 2009, 20:41

QUOTE (shadowking @ Aug 3 2009, 07:00)

--abr 287 -h

Why exactly 287?

curious,
Michael

/mnt

Aug 4 2009, 00:10

QUOTE (Case @ Aug 3 2009, 20:38)

Just a quick question, are you still using that Sound Blaster Audigy SE quick googling found you mentioned? If so, it would be best if you used 48 kHz resampling when you test.

I using a Sound Blaster X-Fi Xtreme Gamer now, which Creative replaced the Audigy with due to the Audigy's imfamus 48khz resampling issue. Sadly this sound card sounds worse on Vista without using Kernel streaming and WASAPI, yes i know no ABX to prove but i like to know a good way to ABX both software and hardware mixers under Vista.

/mnt

Aug 4 2009, 01:11

I've had a quick go at --abr 287 -h, and so far those tracks seems to be transparent. I will do a retest on ABR tommorow night.

Also i've done a sample of one of the tracks which has 1:00 - 1:30 edited onto wav. Am not sure if the artifact will appear though.

One Time, One Place (Live) [Sample]

This is a "lo-fi" version of our main content. To view the full version with more information, formatting and images, please click here.