Skip to main content

Topic: LAME 3.97a11/12 -V2 --vbr-new regression (Read 12832 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.
  • mixminus1
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Here's the clip:

http://www.hydrogenaudio.org/forums/index....showtopic=37002

3.96.1 adds a very slight "coarseness" to the sound.  3.97a11 and a12 add a coarseness that is...well, not so slight...

Edit: I think it's pretty obvious, but just to clarify:  The coarseness only occurs on the first note that is sung...the rest of the clip sounds fine in all versions of LAME (at least to me).  I posted the entire musical phrase so you would know it is a human voice, otherwise it could almost be mistaken for some kind of weird test tone. 
  • Last Edit: 11 September, 2005, 02:23:02 AM by mixminus1
"Not sure what the question is, but the answer is probably no."

  • evereux
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #1
Wow, the noise is pretty bad there on 3.97 a12. It's still audible (as you've said) with 3.96.1 --vbr-new and also 3.97 a12 -V2.

This is an interesting sample. Thanks.
daefeatures.co.uk

  • guruboolez
  • [*][*][*][*][*]
  • Members (Donating)
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #2
I confirm the noise issue.
-V2 --vbr-new has serious problems here.
-V2 has the same noise problem, but at much lower intensity.
3.90.3 and 3.96.1 --standard [not fast] are not clean either.

BTW, 128 cbr/abr encodings suffer from the same kind of warbling with 3.97a (it's more pronounced IMO on the second half of this sample). But it's the first time I see this problem also occuring with VBR

To note: very low bitrate at -V2 --vbr-new during the incriminated part.

Nice sample

  • tycho
  • [*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #3
Is it only lame that struggles with this one? Similar for Vorbis, MPC?
(I know, I should test myself, but right now I don't have my usual soundcard/ earbuds).

/Edit: the above is maybe OT, but if the problem doesn't exist in other encoders, I guess it is more likely that it can be fixed in lame too.
  • Last Edit: 12 September, 2005, 08:09:03 AM by tycho

  • Squeller
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #4
Interesting sample, I can here this very easily even with my cheap pc speakers.

Tycho: No, in current oggenc2.exe @ -q6 it's not there.

  • Donunus
  • [*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #5
how do i convert this flac file to mp3 using eac??? it complains saying its not a wav file

  • beto
  • [*][*][*][*][*]
  • Members (Donating)
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #6
EAC was not designed for that. Decode it with the flac decoder and encode the resulting wav with LAME.
Other option would be using a version of LAME that accepts flac files but i don't recall if that exists....

edit: grammar
  • Last Edit: 16 September, 2005, 03:05:32 PM by beto

  • JunkieXL
  • [*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #7
foobar2k is the easiest way to convert flac in my opinion. Just configure the diskwriter and copy over the lame.exe version you like into your fb2k folder. Plus you can store various command line options and access all them through a drop down list.  Pretty cool if you ask me.  And it supports a huge amount of formats to transcode to and from.
J
  • Last Edit: 16 September, 2005, 02:48:21 PM by JunkieXL

  • Alex B
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #8
Hi,

This is my first post here, but I tried to do this properly.

I decided to make a few high bitrate encodings with several different encoders after hearing how obvious the artifact is with LAME 3.97b1 -V 2 --vbr-new.

I chose: LAME 3.90.3, LAME 3.97b1, Musepack 115v alpha, Vorbis 1.1.1 aoTuvB4 and iTunes 5.0.0.35 AAC.

Here are my ABX results:

LAME 3.97b1 -V 2 --vbr-new (158 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:47:10

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 2 --vbr-new.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:47:11 : Test started.
14:50:55 : 01/01  50.0%
14:51:10 : 02/02  25.0%
14:51:28 : 03/03  12.5%
14:51:37 : 04/04  6.3%
14:51:49 : 05/05  3.1%
14:51:54 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel.


LAME 3.97b1 -V 2 --vbr-new -b 128 (165 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:52:23

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 2 --vbr-new -b 128.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:52:25 : Test started.
14:52:49 : 01/01  50.0%
14:52:53 : 02/02  25.0%
14:53:00 : 03/03  12.5%
14:53:12 : 04/04  6.3%
14:53:21 : 05/05  3.1%
14:53:23 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Not better than without the -b 128 switch.


LAME 3.97b1 -V 0 --vbr-new (188 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:52:23

foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:59:02

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 0 --vbr-new.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:59:03 : Test started.
14:59:17 : 01/01  50.0%
14:59:22 : 02/02  25.0%
14:59:29 : 03/03  12.5%
14:59:36 : 04/04  6.3%
14:59:43 : 05/05  3.1%
14:59:53 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). Not better than -V 2 --vbr-new


LAME 3.97b1 -V 0 --vbr-new -b 128 (192 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:00:28

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 0 --vbr-new -b 128.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

15:00:29 : Test started.
15:00:42 : 01/01  50.0%
15:00:46 : 02/02  25.0%
15:00:51 : 03/03  12.5%
15:00:55 : 04/04  6.3%
15:01:00 : 05/05  3.1%
15:01:02 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Not better than -V 0 --vbr-new without the -b 128 switch.


LAME 3.97b1 -V 2 (161 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:54:02

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 2.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:54:03 : Test started.
14:54:29 : 01/01  50.0%
14:54:40 : 02/02  25.0%
14:55:04 : 03/03  12.5%
14:55:16 : 04/04  6.3%
14:55:24 : 05/05  3.1%
14:55:27 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Much better than any 3.97b1 --vbr-new, but still easy to ABX.


LAME 3.97b1 -V 0 (188 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:56:43

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 -V 0.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:56:44 : Test started.
14:57:08 : 01/01  50.0%
14:57:31 : 02/02  25.0%
14:57:53 : 03/03  12.5%
14:58:17 : 04/04  6.3%
14:58:33 : 05/05  3.1%
14:58:35 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Probably not any better than 397.b1 -V 2.


LAME 3.90.3 --alt-preset fast standard (160kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:52:23

foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:04:01

File A: file://D:\Test\herding_calls sample\herding_calls 3903 --alt-preset fast standard.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

15:04:02 : Test started.
15:04:15 : 01/01  50.0%
15:04:22 : 02/02  25.0%
15:04:30 : 03/03  12.5%
15:04:38 : 04/04  6.3%
15:04:45 : 05/05  3.1%
15:04:47 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel.


LAME 3.90.3 --alt-preset fast extreme (180 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:52:23

foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:04:01

File A: file://D:\Test\herding_calls sample\herding_calls 3903 --alt-preset fast standard.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

15:04:02 : Test started.
15:04:15 : 01/01  50.0%
15:04:22 : 02/02  25.0%
15:04:30 : 03/03  12.5%
15:04:38 : 04/04  6.3%
15:04:45 : 05/05  3.1%
15:04:47 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Not better than 3.90.3 --alt-preset fast standard.


LAME 3.90.3 --alt-preset standard (160 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:02:19

File A: file://D:\Test\herding_calls sample\herding_calls 3903 --alt-preset standard.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

15:02:20 : Test started.
15:02:36 : 01/01  50.0%
15:02:44 : 02/02  25.0%
15:02:52 : 03/03  12.5%
15:03:00 : 04/04  6.3%
15:03:05 : 05/05  3.1%
15:03:06 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Much better than 3.90.3 --alt-preset fast standard & extreme.


LAME 3.90.3 --alt-preset extreme (197 kbps) vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:02:19

foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 15:05:05

File A: file://D:\Test\herding_calls sample\herding_calls 3903 --alt-preset extreme.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

15:05:06 : Test started.
15:05:27 : 01/01  50.0%
15:05:35 : 02/02  25.0%
15:05:43 : 03/03  12.5%
15:05:51 : 04/04  6.3%
15:05:57 : 05/05  3.1%
15:06:01 : Test finished.

 ----------
Total: 5/5 (3.1%)
- a low pitched rattling artifact during the first part (0 - 3.5s). More pronounced in the left channel. Probably not any better than 3.90.3 --alt-preset standard.


LAME 3.97b1 --preset CBR 160 vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:16:21

File A: file://D:\Test\herding_calls sample\herding_calls 397b1 --preset cbr 160.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:16:22 : Test started.
14:17:26 : 01/01  50.0%
14:18:56 : 02/02  25.0%
14:20:50 : 03/03  12.5%
14:21:11 : 04/04  6.3%
14:21:30 : 05/05  3.1%
14:24:29 : Test finished.

 ----------
Total: 5/5 (3.1%)
- distortion at the beginning of the second part (at 3.5 s).


LAME 3.90.3 --preset CBR 160 vs. original
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2005/09/17 14:32:39

File A: file://D:\Test\herding_calls sample\herding_calls 3903 --alt-preset cbr 160.mp3
File B: file://D:\Test\herding_calls sample\herding_calls.flac

14:32:40 : Test started.
14:36:48 : 01/01  50.0%
14:37:26 : 02/02  25.0%
14:38:11 : 03/03  12.5%
14:38:33 : 04/04  6.3%
14:39:00 : 05/05  3.1%
14:39:05 : Test finished.

 ----------
Total: 5/5 (3.1%)
- distortion at the beginning of the second part (at 3.5 s), but slightly better than 3.97b1.


LAME 3.97b1 -V 2 (161 kbps) vs. LAME 3.90.3 --alt-preset standard (160 kbps)
- couldn't ABX

LAME 3.97b1 -V 2 new (158 kbps) vs. LAME 3.90.3 --alt-preset fast standard (160 kbps)
 - couldn't ABX

Also, I couldn't ABX these vs. the original:

LAME 3.90.3  --alt-preset  cbr 192
LAME 3.90.3  --alt-preset insane
LAME 3.97b1  --preset cbr 192
LAME 3.97b1  --preset insane

Musepack 115v  radio (146 kbps)
Musepack 115v  standard (198 kbps)
Musepack 115v  extreme  (240 kbps)

Vorbis 1.1.1 aoTuVB4  Q5  (150 kbps)
Vorbis 1.1.1 aoTuVB4  Q6  (168 kbps)
Vorbis 1.1.1 aoTuVB4  Q7  (188 kbps)
Vorbis 1.1.1 aoTuVB4  Q8  (212 kbps)

iTunes 5.0.0.35 AAC  CBR 160
iTunes 5.0.0.35 AAC  CBR 256
iTunes 5.0.0.35 AAC  VBR 160 (160 kbps)
iTunes 5.0.0.35 AAC  VBR 256 (256 kbps)


It seems the problem is not specific to 3.97 version. The same artifact is present in all LAME VBR samples and it is more pronounced with the Fast switch. Versions 3.90.3 and 3.97b1 are quite similar.

A different artifact is introduced with 160 kbps LAME CBR samples. Increasing bitrate to 192 kbps seems to help.

In my opinion the three other encoders were perfect with this sample.


I used the Terratec DMX6 fire 24/96 soundcard and KOSS PortaPro headphones. I also confirmed some of the results with my old, but still excellent KOSS HV/1A headphones.

The encoded files are available here:  herding_calls_sample.zip (I can keep them on my server for a week or so.)
REMOVED, please make your own encodings.


[span style=\'font-size:7pt;line-height:100%\']Edit: The bitrate for Musepack Extreme was wrong.[/span]

[span style=\'font-size:7pt;line-height:100%\']Edit2: Removed the sample package.[/span]
  • Last Edit: 02 December, 2005, 10:47:00 AM by Alex B

  • beto
  • [*][*][*][*][*]
  • Members (Donating)
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #9
Thank you for sharing your results 

Next time, please ABX successfully at 8 trials at least to get more meaningful results.

If you are interested I can host the samples for you, just send me a PM.

thanks again.

  • moozooh
  • [*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #10
Well, not to support the wrong ways, but I think it's pretty enough when the difference is obvious.
Alex B, thank you for your effort.
Infrasonic Quartet + Sennheiser HD650 + Microlab Solo 2 mk3. 

  • Alex B
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #11
Yes, the differences were indeed obvious. I could have just shortly confirmed the guruboolez's findings, but I think I found some things that the LAME developers might find useful:

1. The LAME CBR 160 kbps and 192 kbps samples don't have the first artifact (0 - 3.5s).

2. Setting a minimum bitrate with 3.97 VBR doesn't help.

Actually, after posting the results I tried also LAME 3.90.3 and 3.97b1 preset extreme & fast extreme with -b 192. None of the four samples were transparent. This time I ABXed them 8/8. 397b1 -V 0 -b 192 was the toughest, but it was the last in the line and it's possible that I just got tired.

3. It is strange that this LAME VBR sound artifact seems to be coming from the left hand side. I tried the both headphones reversed just to make sure that I don't have hearing problems. I didn't.

[span style='font-size:8pt;line-height:100%']Edit: typo[/span]
  • Last Edit: 17 September, 2005, 02:58:19 PM by Alex B

  • user
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #12
2. Setting a minimum bitrate with 3.97 VBR doesn't help.

Imo, iirc, setting high  minimum bitrates manually by switches to VBR modes don't help anything to add more bitrate to certain music parts, where the VBR encoder thinks to add less bitrate.
The reason: bitrate reservoir, the high min. bitrate offers more bits, but the vbr encoder decides, at this frame i don#t need so much bits, so I take unused bits to next frames, and so forth. So, you won#t change much  the encoding results of the vbr encoder.
www.High-Quality.ch.vu -- High Quality Audio Archiving Tutorials

  • beto
  • [*][*][*][*][*]
  • Members (Donating)
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #13
Quote
Well, not to support the wrong ways, but I think it's pretty enough when the difference is obvious.
Alex B, thank you for your effort.
[a href="index.php?act=findpost&pid=327582"][{POST_SNAPBACK}][/a]


Well, we have to draw the line somewhere.... the artifact may not be obvious to everyone.
Maybe I'm just too picky 

  • Alex B
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #14
Quote
2. Setting a minimum bitrate with 3.97 VBR doesn't help.

Imo, iirc, setting high  minimum bitrates manually by switches to VBR modes don't help anything to add more bitrate to certain music parts, where the VBR encoder thinks to add less bitrate.
The reason: bitrate reservoir, the high min. bitrate offers more bits, but the vbr encoder decides, at this frame i don#t need so much bits, so I take unused bits to next frames, and so forth. So, you won#t change much  the encoding results of the vbr encoder.

I added the -b 128 switch to some of the 3.97 samples because the 3.90.3 --alt-presets have a 128 kbps limit internally. I have seen here a few discussions about the removal of the minimun bitrate limit.

You might be right that it doesn't work the same way when the switch is added after the preset. Encspot shows a different bitrate distribution for the whole file after adding -b 128. So it does not just hard limit the lowest bitrate.

In any case, 3.90.3 has the same problem even it uses the 128 kbps minimum bitrate limit internally.


[span style='font-size:8pt;line-height:100%']Edit: typo[/span]
  • Last Edit: 17 September, 2005, 04:37:59 PM by Alex B

  • Alex B
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #15
Quote
Thank you for sharing your results  

Next time, please ABX successfully at 8 trials at least to get more meaningful results.

If you are interested I can host the samples for you, just send me a PM.

thanks again.[a href="index.php?act=findpost&pid=327578"][{POST_SNAPBACK}][/a]

Thanks. As I already said, the artifacts were obvious and I could have gone through each test 100 times with the same results (100/100). Would it have been enough to mention that or should I generally always do 8 trials or more?

I have a limited web space. If you think it would be useful to share the zip package longer, please go on and PM me the link. I'll add it to the post. I wanted to share the 27 lossy samples because I though it would make it easy for others to quickly try them without going through the encoding and file naming stage.

  • shadowking
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #16
Since it was the LAME encoder 12 x 5 trials = 60/60 is more than enough. However if you tried a different encoder or wanted to do only one trial then you aim for <1% pval on that 1st trial.
wavpack -b4x4s1c

  • mixminus1
  • [*][*][*][*][*]
LAME 3.97a11/12 -V2 --vbr-new regression
Reply #17
Was doing some cbr testing of 3.97b1 and came across an interesting discovery with this sample (herding_calls).

As both guruboolez and AlexB have noted, the distortion is worse on the second half of the sample with cbr...at the default -q setting (3).  Using -b 128 -q 2 in 3.97b1 (and a12)  introduces a distortion on the first part of the sample that is...quite extraordinary.  -q 1 lessens it somewhat, but it's still there.  Also, the same effect can be heard with -b 160 -q 2...at 192, it is greatly reduced.

Neither 3.96.1 nor 3.90.3 exhibit this behavior, i.e. -q 2 or -q 1 do not significantly worsen the distortion.

While this may raise the issue of -h (which still maps to -q 2) as the "high(er) quality" switch (certainly, it's only one sample, but it definitely doesn't improve things...), my hope is that it may give some insight into what is causing the distortion with -V2 --vbr-new, seeing as it appears to be the same distortion, but at a much greater level.
"Not sure what the question is, but the answer is probably no."