From the recent Xiph IRC meeting:



Hopefully the wait won't be too long. IIRC 1.0.2 is apparently the 20031230 version of libvorbis so 1.0.3 (aka 1.1) shouldn't be too far around the corner, assuming Monty sticks to his 'frequent release' policy.

QK,

many thanks for the update....this is very good and welcome news.

rs

Hope this will fix different noise/HF issues with vorbis

That is certainly what we'll be expecting as noise normalization was one of the major changes between RC3 and 1.0. Also, if the stereo mechanism of Vorbis will be improved, then I think most of the HF issues will be fixed. Let's just pray that Monty gets it right.

For everybody not knowing where this information comes from:
http://www.xiph.org/minutes/2004/april/raw/

Noise normalization is the technique used by the low bit rate. At an official encoder (1.0/1.0.1), it is used less than q5. I don't consider that this is the main causes of the noise/HF problem. However, it has a certain kind of peculiarity.



I have a different view again. The answer is in aoTuV experiment. This is not converting a stereo at all. And it has an effect also in a monophonic sample.
Probably, a noise/HF problem is complex.

[edit]
typo fix

Do you think Monty and the folks at Xiph review the code of some of the recent 'alternate versions' (GT3b2, aoTuV, etc?)

I wonder if these will be used or if Xiph has some new algorithm in mind. I would not think they would 'reinvent the wheel' if one of these versions is deemed 'best.'

Just curious...

Xiph's attitude seems to be "we creare reference encoders." If I remember correctly, they kicked assembly optimizations out of Theora cvs because it would have made the source code not-so-understandable. They'll propably leave "hard" tweaking for others.

The current NN code only restores energy in case of energy loss (due to quantization). This is a rare case for very low SNRs. Actually most of the times energy will be increased by quantization and Monty was not aware of this. So, in future encoder versions an enhanced NN algo will probably be active for *all* quality levels to compensate for both situations.
This quantization issue is currently the only reason for the HF boost effect for monaural or q>=6 files. However it *only* results in a boost of roughly 1 dB at worst which is hard to perceive.
The point stereo thing seems to have a much greater impact. I'm curious about what Monty meant by "generalizing stereo coupling" - We'll see...

bye,
Sebastian

"Generalizing stereo coupling" means that channel coupling will work with more than two channels.

Awesome. I'm not going to miss Dolby Surround/prologic one bit.