Hydrogenaudio Forums > lame -V 5 --vbr-new extremly problematic sample

Full Version: lame -V 5 --vbr-new extremly problematic sample

Hydrogenaudio Forums > Hydrogenaudio Forum > Uploads

gasmann

Jan 19 2008, 15:10

Hi folks,

I've incidentially found an obviously very hard to encode song. I don't know why, but it sounds extremely bad when encoded with lame 3.97 -V 5 --vbr-new. lame 3.98b6 is a lot better (at same setting of course) but still absolutely easy to abx. I'm normally fine with V 5, but this one is really bad. And I mean, it's the beginning of a normal song, not a killer sample (!!) and it feels really unlistenable to me

Please have a look at this, maybe it can help to further improve lame

raintheory

Jan 19 2008, 15:33

Harpsichords are pretty difficult to get a decent mp3 from... I bet that is the culprit here.

Bodhi

Jan 19 2008, 16:28

Haaaa, Final Fantasy...

Not a single problem to me with my tools (including my ears

). It doesn't annoy me AT ALL!

I understand now why I'm not the kinda guy who makes the difference between a CD and V2 (or even V5 apparently

).

The good thing is that it allows me to keep plenty of place on my hard disk

So, sorry, I'm helpless.

I'm out

retro83

Jan 19 2008, 16:50

I can also ABX that at V5 with no problems. Wonder why harpsichords are hard to compress?

Sunhillow

Jan 19 2008, 18:00

Harpsichords produce a rich spectrum of audible harmonics. So the encoder tries to treat them all but does not have enough bits to do so.

FYI, harpsichord tunes were already used to test MUSICAM and ASPEC, the predecessors of MP2 and MP3

Edit: this is not a harpsichord, they sound even more complex. Is it an acoustic guitar played with a plectrum?

naturfreak

Jan 19 2008, 18:12

I'm not sure that instrument is really a harpsichord. To my ears it sounds more like a highpass filtered guitar with strings made of steel. There are fret noises too.

The bitrate is quite low for -V 5 --vbr-new setting.

soultrain

Jan 19 2008, 18:28

QUOTE(naturfreak @ Jan 20 2008, 01:12)

Why not use -V2 that should sound much better, have you tried it. V5 is not very much to work with.

naturfreak

Jan 19 2008, 18:57

I consider myself as untrained in ABXing, because I did last ABX test is several mounths ago.

I just ABXed LAME 3.98b6 -V 3 --vbr-new vs. lamekiller original successfully. I'll try -V2 setting later.

QUOTE

foo_abx 1.3.1 report
foobar2000 v0.9.4.4
2008/01/20 01:41:44

File A: F:\lamekiller.wav
File B: F:\lamekiller.mp3

01:41:44 : Test started.
01:43:07 : 01/01 50.0%
01:43:39 : 02/02 25.0%
01:44:10 : 03/03 12.5%
01:44:31 : 04/04 6.3%
01:45:17 : 05/05 3.1%
01:45:44 : 06/06 1.6%
01:46:13 : 07/07 0.8%
01:46:35 : 08/08 0.4%
01:47:04 : 09/09 0.2%
01:47:26 : 10/10 0.1%
01:47:50 : 11/11 0.0%
01:47:58 : Test finished.

----------
Total: 11/11 (0.0%)

Irakli

Jan 19 2008, 19:33

Yes, it very easy to ABX, even for an extremely untrained listener like me (I usually don't do abx and just encode everything to -V2).

Here's my result.

lame3.97 -V 5 --vbr-new

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 00:58:25

File A: C:\Documents and Settings\Иракли\Desktop\lamekiller.flac
File B: C:\Documents and Settings\Иракли\Desktop\lamekiller.mp3

00:58:25 : Test started.
01:01:52 : 01/01  50.0%
01:02:09 : 02/02  25.0%
01:02:33 : 03/03  12.5%
01:03:15 : 04/04  6.3%
01:04:00 : 05/05  3.1%
01:04:35 : 06/06  1.6%
01:05:15 : 07/07  0.8%
01:05:41 : 08/08  0.4%
01:07:11 : 09/09  0.2%
01:07:56 : 10/10  0.1%
01:08:07 : Test finished.

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

lame3.98beta6 -V 5 --vbr-new

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 01:18:11

File A: C:\Documents and Settings\Иракли\Desktop\lamekiller.flac
File B: C:\Documents and Settings\Иракли\Desktop\lamekiller.mp3

01:18:11 : Test started.
01:18:28 : 01/01  50.0%
01:18:38 : 02/02  25.0%
01:18:48 : 03/03  12.5%
01:19:01 : 04/04  6.3%
01:19:08 : 05/05  3.1%
01:19:20 : 06/06  1.6%
01:19:45 : 07/07  0.8%
01:19:51 : 08/08  0.4%
01:20:06 : 09/09  0.2%
01:20:25 : 10/10  0.1%
01:20:27 : Test finished.

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

I will try to test with -V 2 soon.

Best wishes,
Irakli

shadowking

Jan 19 2008, 20:07

Its sounds like a tremolo effect similar to my angels fall first sample but not nearly as bad. I tried my linux box that has lame 3.96.1. V5 is obvious, but with V3 i wouldn't notice it on a normal listen.

3.98b6 sounds a little worse but i didn't abx. V2 is about the same as V3 - a little off on first notes. V5 is obvious. Will try abx when I have time.

soultrain

Jan 20 2008, 04:38

V2 should be good enough for most (if not all) encodings, but i prefer to use v1 for when there is a problem sound, maybe its not needed, but it give me a better feeling, to have some extra bits free to work with.

halb27

Jan 20 2008, 05:13

That's why I use even -V0. Before 3.98b6 (maybe already some versions earlier) there wasn't much use IMO to prefer it over -V1, but now when these rare problems come up sometimes -V0 gets a noticable better quality. I get an average bitrate of 235 kbps for my music which I can easily accept.
Remaining problems with -V0 have never been annoying to me.

gasmann

Jan 20 2008, 09:36

Someone here suggested going to -V 2 --vbr-new. Apart from feeling uncomfortable with such big mp3s because I use them mainly on my old 256mb flash player, it seemed a really good choice. Well, it seemed. Both 3.97 and 3.98b6 were easily ABXable on -V 2 my me. Of course, it feels somewhat better, but it's far from being acceptable

I also second what naturfreak said; the bitrate is really low for -V 2! 3.97: 138kbps, 3.98b6: 152kbps

Given that 3.98b6 generates more "bloated" files in general, I think both behave the same in this regard. The encoder obviously doesn't realize it's hard to encode. Or it sees that more bits are just wasted because it doesn't sound much better anyway (V2 isn't much better than V5 to me)

3.97:

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 16:21:33

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv2_397.mp3

16:21:33 : Test started.
16:21:40 : 01/01  50.0%
16:21:44 : 02/02  25.0%
16:21:46 : 03/03  12.5%
16:21:50 : 04/04  6.3%
16:21:55 : 05/05  3.1%
16:21:59 : 06/06  1.6%
16:22:22 : 07/07  0.8%
16:22:24 : Test finished.

----------
Total: 7/7 (0.8%)

3.98b6:

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 16:23:42

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv2_398.mp3

16:23:42 : Test started.
16:23:46 : 01/01  50.0%
16:23:51 : 02/02  25.0%
16:23:54 : 03/03  12.5%
16:23:58 : 04/04  6.3%
16:24:02 : 05/05  3.1%
16:24:05 : 06/06  1.6%
16:24:16 : 07/07  0.8%
16:24:17 : Test finished.

----------
Total: 7/7 (0.8%)

soultrain

Jan 20 2008, 10:17

Gasman: thansk for testing, good to know there are still problems even at V2.
I see your problem with big files on a 256mb player :-) If you have the time could you test it also with V1 and V0? i am curious or the problem is solvable with enough bits or not.

gasmann

Jan 20 2008, 12:30

Whew, my first listening test ever on lame -v 0 --vbr-new was successful

On 3.97, it was not a big deal. But on 3.98b6, I really had a hard time to hear what's the mp3 and what's the flac. Even with 3.97, I'd say it's well above 'OK' for casual listening. So you're right, halb27

Just for your information, the vorbis encoder aotuv b5 faithfully encodes this thingy, so that I can't (or hardly , my ears are somewhat stressed now) hear any difference at q4. Even q2 seems kinda OK, about as good as 3.97 -V0. And that at half of the bitrate!

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 18:55:46

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv0_397.mp3

18:55:46 : Test started.
18:55:57 : 01/01  50.0%
18:56:02 : 02/02  25.0%
18:56:06 : 03/03  12.5%
18:56:32 : 04/04  6.3%
18:56:40 : 05/05  3.1%
18:56:51 : 06/06  1.6%
18:56:59 : 07/07  0.8%
18:57:00 : Test finished.

----------
Total: 7/7 (0.8%)

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 19:00:06

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv0_398b6.mp3

19:00:06 : Test started.
19:00:26 : 01/01  50.0%
19:02:01 : 01/02  75.0%
19:02:10 : 02/03  50.0%
19:02:17 : 02/04  68.8%
19:02:26 : 02/05  81.3%
19:02:33 : 03/06  65.6%
19:02:49 : 04/07  50.0%
19:02:57 : 05/08  36.3%
19:03:08 : 06/09  25.4%
19:03:38 : 07/10  17.2%
19:03:41 : 08/11  11.3%
19:03:44 : 09/12  7.3%
19:03:47 : 09/13  13.3%
19:03:52 : 10/14  9.0%
19:04:00 : 10/15  15.1%
19:04:09 : 11/16  10.5%
19:04:31 : 12/17  7.2%
19:04:42 : 13/18  4.8%
19:05:57 : 14/19  3.2%
19:06:09 : 15/20  2.1%
19:06:16 : 16/21  1.3%
19:06:24 : 16/22  2.6%
19:06:32 : 17/23  1.7%
19:06:55 : 17/24  3.2%
19:07:38 : 17/25  5.4%
19:08:13 : 17/26  8.4%
19:08:30 : 18/27  6.1%
19:11:03 : 19/28  4.4%
19:11:25 : 19/29  6.8%
19:11:56 : 20/30  4.9%
19:12:03 : 21/31  3.5%
19:12:09 : 22/32  2.5%
19:12:15 : 23/33  1.8%
19:12:21 : 24/34  1.2%
19:12:30 : 24/35  2.0%
19:12:45 : 25/36  1.4%
19:13:31 : 25/37  2.4%
19:13:40 : 26/38  1.7%
19:13:48 : 27/39  1.2%
19:13:56 : 28/40  0.8%
19:13:58 : Test finished.

----------
Total: 28/40 (0.8%)

soultrain

Jan 20 2008, 13:10

Nice to see that the problem is gone

its a pity that most mp3 player dont handle vorbis

, ~~better quality~~ the same quality at half the bitrate sounds good.

Could you also test the V1 to see how close (or far) it is to V0?
Thanks

gasmann

Jan 20 2008, 13:35

Test with V 1 shows that it's OK, too. I'd says it's only slightly worse than V 0. I didn't have much trouble ABXing it, but really, especially with 3.98b6 it's fine for normal listening (a lot better than V 2 at least). So, V 1 is probably the best compromise between filesize and quality for me (for this particular track).

However, having to mess around with V settings just to find the right one for each song is kinda time-wasting. I wish there was a "super-VBR" mode in lame that could automatically detect what bitrate range is best for the given sample. A VBR mode even more variable than the current ones, resulting in even less predictable bitrates. Something that can go as high as 256kbits avg. or as low as 64kbits avg. depending on the complexity would be cool

Anyway, I didn't say it sounds better at half of the bitrate, just about as good. That's a big difference, at least to me.

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 20:16:44

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv1_397.mp3

20:16:44 : Test started.
20:17:01 : 01/01  50.0%
20:17:08 : 02/02  25.0%
20:17:18 : 03/03  12.5%
20:17:25 : 04/04  6.3%
20:17:34 : 05/05  3.1%
20:22:31 : 06/06  1.6%
20:22:35 : 07/07  0.8%
20:22:36 : Test finished.

----------
Total: 7/7 (0.8%)

CODE

foo_abx 1.3.1 report
foobar2000 v0.9.5
2008/01/20 20:22:49

File A: L:\Eigene Dateien\Downloads\lamekiller.flac
File B: O:\_ lamekillerv1_398b6.mp3

20:22:49 : Test started.
20:23:00 : 01/01  50.0%
20:23:06 : 02/02  25.0%
20:23:13 : 03/03  12.5%
20:23:19 : 04/04  6.3%
20:25:00 : 05/05  3.1%
20:25:06 : 06/06  1.6%
20:25:12 : 06/07  6.3%
20:25:26 : 07/08  3.5%
20:25:33 : 08/09  2.0%
20:25:49 : 09/10  1.1%
20:25:58 : 10/11  0.6%
20:25:59 : Test finished.

----------
Total: 10/11 (0.6%)

halb27

Jan 21 2008, 02:19

QUOTE(gasmann @ Jan 20 2008, 21:35)

However, having to mess around with V settings just to find the right one for each song is kinda time-wasting. I wish there was a "super-VBR" mode in lame that could automatically detect what bitrate range is best for the given sample. A VBR mode even more variable than the current ones, resulting in even less predictable bitrates. Something that can go as high as 256kbits avg. or as low as 64kbits avg. depending on the complexity would be cool

VBR does what you want, it's just not perfect.
With mp3 it's always like this: with a quality setting that corresponds to ~130 kbps everything's fine usually, but there are problem areas where this is not so. Usually you have to use a quality setting that corresponds to 200+ kbps in order to get the problem transparent or at least not annoying.
In principle this is the same for Vorbis or AAC, but the probability of getting at an annoying problem is significantly lower at ~130 kbps, and already at ~160 kbps the degree of security against bad encodings is close to 0.
If I had a 256 MB DAP that supports just mp3 (and probably WMA) I'd use 3.98b6 -V5 usually and switch to a higher quality setting in those cases that are annoying to me with -V5.

Francksoy-Wizzu

Jan 21 2008, 08:22

QUOTE(gasmann @ Jan 20 2008, 19:30)

Whew, my first listening test ever on lame -v 0 --vbr-new was successful

Mmmh. Looks like we have similar ears and expectations.

Totally obvious tremolo for me up to V3, @ V2 and V1 I can still ABX but I must concentrate.. you have better ears than me because couldn't ABX @V0.

Transparent to my ears with Vorbis AoTuV b5 / V3.

I'm not anywhere near to get back to MP3 anytime soon...

BTW I'm pretty sure there is no acoustic instrument playing at all in this track, it's 100% samples played from a keyboard, I'd put my hands to the stake on it. The samples sound like they're from extremely pristine, metallic guitar strings, artificially (digitally) pitched upwards. The "squeek" sound at the beginning is probably a practical joke so to make believe there actually is a guitar playing.

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