Source:

http://sjeng.org/ftp/vorbis/libvorbis-1.0gt3b2.tar.gz

This is based off RC3+patches. I did this because there are some issues with 1.0, that originated between RC3 and 1.0, and that I haven't been able to track down yet. Because it's based on RC3 it's lacking some stuff that's in 1.0, but that doesn't seem to be relevant for high bitrates. I can't guarantee there won't be any regressions compared to 1.0 or GT3b1 though, but it should improve the (HF) noise issues people have been having.

It's tuned from -q4 to -q10.

Edit: And don't try anything but 44.1/48kHz stereo either. And test it before you encode your entire collection, I have no idea how good it works.

Because it's based on RC3 and not 1.0, it is not going to be plug-in compatible in applications. You'll need to make adjustments to the encoder tools (back to RC3 interface) to get it to work. This is major pain in the ass, which is why I didn't upload an executable, I'll let someone else cope with that.

Obviously, I'll want to get back to 1.0 for the next beta, but I want to see what results this gives, and I've worked on it anyway and I know people like to toy with a new version, so there you go.

Here's win32 binary of oggenc.

Thank you, Case.
I've just tried on an easy but unvalid sample (short_block_test_2), in order to check the behaviour for this new GT3 build on pure and ideal attacks. In this special conditions, GT3b2 is clearly worse than the beta1 (very good, this one).

GT3b1 -q5 = 485 kbps
GT3b2 -q5 = 327 kbps

I need to try on valid samples, like castanets.wav or others passage.

Transient behaviour in general is going to be a bit less responsive than beta 1, although with noticeable bitrate savings and, I think, no real disadvantage for normal music.

I tried on the easiest sample after short_blocks : castanets2.wav



Include : a post-CVS encoding at -q5 (oggdropXPD 1.6.11c).

Progress between GT3b2 and Post-CVS : less pre-echo, but too easy to ear, and really annoying. GT3b1 is obviously the best, though I can ABX it without real difficulties. Intersesting to note : bitrate is smaller with beta1, for a much higher quality.

PostCVS = 166 kbps
oggGT3b1 = 197 kbps
oggGT3b2 = 205 kbps

Gah, that's bad news. No good then.

Keep working Garf, all true Ogg Vorbis zealots (like me) are with you!

I tried now on a short sample (early but not too sharp percussions).

oggenc PostCVS -q 5 : 154 kbps
oggenc GT3b1 -q 5 : 162 kbps
oggenc GT3b2 -q 5 : 175 kbps



On ABA test, it wasn't too difficult to hear a difference, but really difficult to differenciate the 3 encoding. With ABX, the third sample (GT3b2) was much harder to perform (good beginning [8/12], ruined by a bad ending). Less noise with this sample.

P.S. I had to put the volume on a excessive level in order to perform this test.

[COLOR=red]Sorry for flooding

Tried on harpsichord.

oggenc PostCVS -q 5 : 201 kbps
oggenc GT3b1 -q 5 : 216 kbps
oggenc GT3b2 -q 5 : 208 kbps



On ABA test, postCVS was immediatly perceptible (kind of pre-echo, or distorsion). Can't distinguish GT3 series.
With ABX, all were abxable, especially on noise (though I focused on distorsion for the last one, GT3b1). I obtained the best ABX result with GT3b2. And again, GT3b2 was the biggest file [EDIT : not true] :-/

Did you find that GT3b2 had less or more noise than GT3b1?

Well, it's not easy to say (and I have now other samples in memory...). On harpsichord, there's a kind of mixture between pre-echo on the moment of attack, and the development of the note, noisy and sometimes distorted.

Other sample : JUMP, 30 seconds. I listened three passages :

1/ the electric introduction (first three seconds)
2/ short-spaced drums (13.3-14.5)
3/ cymbals (14.5-15.0)

1 · INTRODUCTION


beta1 and CVS are really close, and they share the same flaw (short impulsion of noise after each guitar (?) note). GT3b3 is near perfection (as RC3)

2 · DRUMS :



Again, GT3b2 is the best. As with harpsochord, I can't analyse and describe the exact flaw of CVS/GT3b1 : is it pre-echo, or noise ? Attacks sound different, and a bit annoying to me.
Gt3b2 is not transarent, but need more concentration for ABXing (I performed quick tests).


3 · CYMBAL



Identity between Post-CVS and GT3b1 : same flaw on cymbal and similar sound. The micro-details are covered by a generated noise. GT3b2 is again near-transparency.



These 30 seconds may be sufficient to proove that final vorbis codec version has a general behaviour that facilitates ABX tests, and make transparency quite impossible at -q5, and in some case, -q 8. RC3, and therefore GT3b2, doesn't suffer by it.
Nevertheless, on strong attacks, GT3b1 is obviously the best vorbis codec I never seen.

Post CVS = 180 kbps
GT3b1 = 203 kbps
GT3b2 = 208 kbps

The bitrate-increase from GT3 beta 1 to beta 2 is really dramatical:

Album "Aerosmith - Big Ones"

total length encoded with b1 at -q6 : 121,616 MB
total length encoded with b2 at -q6 : 137,388 MB = + 12,97 % (!)

And with other albums it´s similar.

I did a short abxing on Castanets.flac. I compared a 1.0 -q6 with a GT3b2 -q5. I just took care of the noise (not preecho or something else).

The result was 10/10

The GT3b2 has less noise than the 1.0.

I'll do another test comparing GT3b1 with GT3b2 and report my results later today.

I haven't really tried on many tracks. I've just encoded one movement of a very recent symphonic recording (Beethoven, fifth Symphony by Sir Simon Rattle, released in 2003) :



+12% between beta1 and beta2.


Noise is considerably reduced. I focused my hearing on the pre-symphonic beginning (noise, chairs, etc...). PostCVS and GT3b1 were immediatly betrayed by a small electric noise in the background.

The two files were in consequence easy to ABX. Nevetheless, the GT3b2 was different too :



The GT3b2 is not neutral at -q5. A small amount of noise, which gave a small coloration is added :


I did only one mistake on the 15 last tests.

Therefore, the GT3b2 is a bit noisy, but this noise is less annoying and perceptible than the CVS based codecs.

There are the anounced results of my second abxing.

castanets.flac
version.................GT3b1.........vs..........GT3b2
bitrate (-q5).......215kbps...................208kbps
encspeed...............6x............................3x

IMHO the huge difference btw. the encoding speed is due to the fact that oggenc only counts entire seconds. In the case of GT3b2 it is slightly slower and just takes a bit more time and counts two seconds.

As I did in the test before I ONLY concentrated on the noise. This time it was harder. I got a:

16/19 --> 84%
Guessing Prob: 0,5%

The GT3b1 has more noise and it (the GT3b1-noise) sounds a bit more coarse (to me )

Keep on tuning Garf! I'd love to see a final GT3 for archieving !

fragtal

EDIT: Damn the not working tabs

I tried to ABX mpc (1.14) --standard. Bitrate is similar on the whole track to GT3b2 -q5 (168 kbps vs 164 kbps).



No noise (for me) with mpc standard, small noise with oggenc GT3b2.

I tried to check higher settings for GT3b2 (Q7 ; 226 kbps) :



I had to put more attention than with Q5 in order perceive a difference. And there is a (very small for me) difference.

I accidentally found the first sample where GT3b2 is really worse than GT3b1, on other things than pre-echo. I hear an obvious and unpleasant distorsion on this harpsichord sample. Especially, between 2.5-5.0 seconds.

Bitrate (whole track) at -q 5 :

Oggenc PostCVS = 201 kbps
Oggenc GT3b1 = 229,2 kbps
Oggenc GT3b2 = 215 kbps


ABX results : 12/12 for beta2 and 6/12 for beta 1.

is it me or does this one seem to encode slower... not that its a problem great work garf

yep, see?

I already postet the same experience.

It's based on older code, so it's indeed quite a bit slower. Based on the results here this seems to be a 'dont use' version anyway.

So Garf, hows the progress on GT3?

Not to be pushing you to another release, but what can we expect for beta3?

I'm not sure what else can be done until the next official release, so will we see a beta3 before the next release on the CVS?

Beta 3 needs 2 things:

1) adaptive way of encoding transients (I worked out the idea, and Dibrom told me APS works similarly, but no code written yet)

2) Pin down noise issue

I have no idea when I will have time for this.

Garf, do your tunings (when they turn out to be good) end up getting merged into the official Vorbis CVS? I'm guessing most people use the official Xiph releases, so if you manage to improve quality, it'd be nice if more than just the people who read HA could take advantage of it.

They do, at least if I release them publicly, which I didn't do with beta 1.

--
never mind

I'm using the official version atm, would it be a good idea to switch to one of these betas? Using -q8

It's definitely worth switching. The official version has not been tuned at settings above about q6, so it starts to get hugely wasteful of bits by q8. Though I don't hink Garf paid a lot of attention to levels quite that high, q5 and q6 were improved, and these improvements should have had a knock-on effect. So it's likely you could use the new version at q7 and not notice any difference, except smaller files!

But I'd suggest something more - if you want to stay with ogg (a fine decision IMHO) then try encoding with the new Garf version at q6 - I suspect you won't notice any difference, except for significantly smaller files. You could probably get away with q5 even, but if you have enough room for the existing q8 then new q6 should be a big enough improvement!

Cheers, Paul