Updated : lame 3.98 beta 6 (small fix) (2008-03-12 13:47)

Sources : http://sourceforge.net/project/showfiles.p...;package_id=309

Binaries : Rare Wares

CVS : http://lame.cvs.sourceforge.net/lame/lame/...me/?sortby=date

Official LAME v3.98 Beta 6 is out on December 16 2007:

• Robert Hegemann:
  • Feature request [ 1811483 ] WAVE_FORMAT_EXTENSIBLE support (PCM)
  • Fix for some rare scalefactor selection issue the newer vbr code had at low compression levels
  • Fix for Bugtracker item [ 1813496 ] AIFF parsing bug
  • Latest changes to the new VBR code:
    • it now has its own psy model, a derivation from NSPSY.
    • some more tuning has been done for this new psy model. Many thanks to Horst Albrecht and Myles Thaiss.
    • the "out-of-bits" strategy is reworked
  • It was possible, that the "interchannel masking effects feature" was used by the dual-channel-mode for bi-lingual encodings too. It was meant to work on stereo L/R channels only.

Changelog - history.html

Can someone test with new killer samples??

For my personal use and maybe if someone wants to test beta 6 out on a Macintosh computer (and isn't command line savvy), I've compiled Max w/ LAME 3.98 beta 6.

http://www.vortec6.com/max071lame398b6.tar.gz

also the command line version in universal binary goodness

http://www.vortec6.com/lame398b6.tar.gz


EDIT: Compiled again for the changes mentioned below

Did a quick test with -V2, for some reason this version produces smaller bitrates then beta 5 and 3.97.



Sounds the same or worse then beta 5.

Here is another killer track I have that is not transparent on LAME 3.90.3 --alt-preset standard and also not transparent on LAME 3.97 -V --vbr-new.



This track was transparent on beta 5, but it not at beta 6 and the beta 6 encode produced a smaller bitrate while ogg at q.6 and nero aac q.55 was in bigger bitrates. So far this the worst encode I heard of this track it also has the artifacts that were on -V 5.

LAME 3.98b5 205kbps
LAME 3.98b6 187kbps
LAME 3.97 187kbps
LAME 3.90.3 192kpbs

worse than b5??
Hmmm......
that's interesting.
anyway...
Thanks~ /mnt

Here is another test I did at V2.



A hard to hear but not annoying warbling at 3:58, but my ears picked it up like there is something wrong.

This encode also produced a smaller bitrate then the past version i tested 192kbps LAME 3.98b6, 211 kbps LAME 3.98b5, 200kbps LAME 3.97 and 220kbps LAME 3.90.3. I find this version be worse then 3.97 and I find beta 5 to be better then 3.90.3 and 3.97.

/mnt, you make me go lossless all the way...
I wonder if you can ABX OGG at q.6 with the EIG sample.
it's so weird these new LAME versions not handling a lot of things...

Codec development is a challenge, and this is especially true for mp3.
The ways towards improvement are not wide and straight.
It's always weighing advantages vs. disadvantages of a certain approach according to experience.
That's why /mnt's findings are so helpful, and why a forum like this is important where members are invited to contribute and share their experiences.

This is a case of selective perception aka ""Hasty Generalization" and "Proof By Example". Both are typical mistakes made by "paranoid" and "idealistic" minds.

A bitrate test on 1 song:

3.98 b5 - V4 vbr new = 159 k
old = 158 k

3.98 b6 - V4 vbr new = 150 k
old = 133 k !!

25 k difference is too sus..

Looks like --vbr-new got recalibrated due to a bug fix.

http://lame.cvs.sourceforge.net/lame/lame/...tize.c?view=log

Its hard to believe that a late beta has such a bitrate gap - 10 k for vbr new and 20k for vbr old. All vbr is different.

Since me and others were very happy about the quality improvements with b5 i hope the devs don´t want to please the few that complaint about a bitrate rise to 3.97.

It looks like a bug or something. vbr old is 30 k lower than B5. I confirm it with another track: 163 k B5 vs. 139 k B6

It seems Robert is working on it:

http://lame.cvs.sourceforge.net/lame/lame/...by=date#dirlist

Thank god I can not ABX EIG (i tried my best) on Ogg at q.0.6 and Nero AAC at q.0.55, but on mp3 its very easy and so for I find iTunes mp3 encoder to do worse on EIG and it did worse then the first verion of l3enc

LAME 3.97 V 2 --vbr-new



I gave up at 8 because it was torture, sounds like a encode from a very old version of l3enc.

LAME 3.98b6 V 2



A nice improvement from 3.97, but still not transparent at 0:11 was louder and still had artifacts awell but almost acpectable.

Thanks , No use testing this one until new builds are out.

New builds at Rarewares now.

John, the LameDrop link for version 3.98b6 seems to be wrong ; it gives me the cli encoder package.

Thanks!

Oooops!! Thanks for the heads-up, I've corrected that now.

You're welcome, thanks for your work ;-)

You did not notice the difference with NERO q 0.55 ?
Which Nero Version did you use?



The dogmatic strange devoted phylosophic writing-like you do reveals deeply to me that you have a strong religion inside your heart. Your religion is SCIENTOLOGY!!! LOL

"a lot of things" meant irony in case you didn't know. But you HAD to take it to the religious path... oh boy....

I used Nero AAC codec / Feb 12 2007 build since the latest version produces a horrible metallic warbling noise on some AAD mixed songs such as Iron Man by Black Sabbath.

http://www.hydrogenaudio.org/forums/index....showtopic=56849

I tested 3.98b6 -V2 with Birds, eig, harp40_1, herding_calls, lead-voice, trumpet, and trumpet_MyPrince.

On the positive side is harp40_1 which was pretty hard to abx for me at -V2.
eig is good too with respect to this extremely hard sample.
The trumpet_MyPrince tremolo problem near the end is also very easy to accept.
Everything else was as expected with the exception of trumpet: second 2.6-5.7 isn't good (also a kind of tremolo) and easily abxable 10/10. I rechecked with 3.98b5: same thing here.
Very strange as I was convinced trumpet wasn't a problem any more due to earlier listening tests with 3.98.I remember having heard it before but either has it turned more serious or may be my listening sensitivity for such a kind of problem has "improved".

Out of curiosity I also tried --vbr-old with these samples. --vbr-old is so clearly worse on nearly all the samples that I suspect there is a bug with --vbr-old. It's very pronounced with lead-voice and herding_calls.

@ /mnt

Well, that is interesting... because I abx'ed Nero-Latest with EIG even at ~300kbps. It's truly a killer codec sample. For OGG, you used aoTuV-latest?

Yeah I used the latest aoTuV ogg encoder at q6 and got 242kbps.

Did someone try the new 3.98b6??
That makes bigger file than old 3.98b6~

I did some quick ABX testing with "Hoobastank - Better" and "Red Hot Chili Peppers - Easily" (which both have been problem samples for my ears as they contain a lot of hi-hat) and while I could always ABX 10/10 with 3.97 -V2 --vbr-new, I haven't been able to get better than 14/20 and 13/20 with 3.98b6 -V2. Keep up the tuning and the good work!

For some reason I was under the impression that there weren't going to be significant changes between 3.98b5 and 3.98 final... and then I saw this. Would you (the collective you) recommend waiting for the final version to do any serious re-encoding?

There is no "official" 3.98 beta 6 yet. I'll let you know when it's done.

Thanks for the quick update!

Phew! I was wondering when someone of the devs was about to react. IMHO, not a good move to test a half-finished release. Especially with the new VBR tunings half-done it caused a lot of havoc.

I think we all have to quickly forget about these ABX results in this thread for now...

I appear to have jumped the gun slightly, here, quite unintentionally. I reacted to the uplift of the version number in 'version.h', which did not seem unreasonable, and I have to question the wisdom of doing that prior to the release being ready.

well until now the release version is usually the last thing to be updated prior to "release" - that's how i've always checked out lame in the past

but i agree, something more official would be ideal


later

what's next after 3.98 ? Lame 4.0 ??

Good question. However, 4.0 has always been and, as far as I know, still is Takehiro's baby and that branch has been inactive for the best part of two years!! So, it's any ones guess, I think. It would be good if he found the time/interest to revive it as it was a radical redesign.

i'm looking forward to 3.100 - that's going to be interesting

no sarcasms accepted... I have seen this 4.0 branch and wonder why what was it...

I tested 3.98b6 using ABR 320 and it yielded 317 kbps on one of my tracks. Previously it yielded 298 kbps on 3.97. What happened here?

http://www.hydrogenaudio.org/forums/index....st&p=524979

I get bigger files with lame 3.98b6
I'm listening mostly to louder music, metal/hardrock and rock influenced

Is 3.97 or 3.98b6 better for metal?

Given some of the results listed in this thread and that the 3.98b6 isn't official, I would wait for the official next beta and see how that compares.

I'm new here so I'm just a little confused - why would there be a Lame 3.98 beta 6 bundle available for download on rarewares if no "official" Lame 3.98 beta 6 is out yet? Would that be like a beta to the beta?

BTW, I'm using the recent 3.98 beta 6 from rarewares and it seems to work pretty well for me - but I'll keep checkign back for the final version.

By way of explanation: The source code is freely available from the LAME project on Sourceforge. The LAME project, per se, does not provide compiled executables for people, such as yourself, to download, people such as I download the latest source code, compile it and offer the resultant executables for people to download and use. In the case of LAME, the version numbering is controlled by one of the library header files, namely - version.h. Historically, this is the last file to be updated as a new release is made. However, in this instance, it would appear that the version number was uplifted in anticipation of the new release. As a consequence I, not unreasonably, I think based upon the history of this, assumed that this was a full, new beta release and announced it as such. On this occasion, apparently the release was not ready!!

There have been some changes since the last compile was made available, but given what has just occurred, I am hesitant to provide any fresh compiles until it is clear that the beta release is complete.

I am finding LAME 3.98 beta to be better for Metal and Hard Rock at V2, I can ABX a few Metal and Hard Rock tracks on LAME 3.97 while I find them transparent on LAME 3.90.3 (preset standard) and LAME 3.98 beta (V2). Also the drums on Metal and Rock is more prone to preecho and LAME 3.98b is less prone to preecho artifacts then LAME 3.97.

Thanks for the explanation and all your hard work John... I'll keep checking rarewares for the new build with the full beta - in the meantime I'll keep using the 3.98 beta 6 that you put up on the 22nd.

I hope we will see final this year

But isn't that a bad idea as the "beta6" was created in the midst of some important (but not yet completed) tuning changes from beta5. Wouldn't it be better to stick to the official beta5 and take down the unofficial beta6 from rarewares as some tunings may be still broken?

Good point... I'd like to know if the unofficial Beta 6 is best to use for -V 0 --vbr-new or if I should be using Beta 5 or even 3.97 final... I'm about to rip a lot of albums and I'd hate to end up redoing it later.

New 3.98b6 came out at rarewares on 2007-11-01.
is that official?

Not as far as I know, it's just the latest CVS build, but I think it fixes one or two things.