IPB

Welcome Guest ( Log In | Register )

 
Reply to this topicStart new topic
GT3b2 with HF Reduction, Discussion
SebastianG
post Mar 23 2004, 16:40
Post #1





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



Split from here.

Please tell me some details about your HF boost reduction strategy.

Maybe my posting at
Ogg Vorbis - General
is of some interest.

edited: had some troubles posting the link ;)

bye,
Sebastian

This post has been edited by dev0: Mar 26 2004, 15:06
Go to the top of the page
+Quote Post
SebastianG
post Mar 24 2004, 00:09
Post #2





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



Accorgint to "the white noise test" i did with your encoder
( see my paper )
you did not seem to fix anything.
Please don't say you've reduced the HF boost effect withoug even knowing the cause.

Correct me, if i'm wrong.

(The things i'm saying in the paper are valid vor q>=6 and somewhat applicable for q<=5)

bye,
Sebastian
Go to the top of the page
+Quote Post
QuantumKnot
post Mar 24 2004, 00:41
Post #3





Group: Developer
Posts: 1245
Joined: 16-December 02
From: Australia
Member No.: 4097



QUOTE (SebastianG @ Mar 24 2004, 09:09 AM)
Accorgint to "the white noise test" i did with your encoder
( see my paper )
you did not seem to fix anything.
Please don't say you've reduced the HF boost effect withoug even knowing the cause.

Correct me, if i'm wrong.

(The things i'm saying in the paper are valid vor q>=6 and somewhat applicable for q<=5)

bye,
Sebastian

I think we are referring to different problems. The HF boost that I and other people have noticed to be reduced, through listening tests on real life music samples, is the excessive brilliance of transients in hi-hats and cymbals. I don't claim to have fixed the source of the problem but I've patched to reduce the effects in cases that are most noticeable. Read the listening tests people have done rather than looking at frequency responses of artifically generated test signals.
Go to the top of the page
+Quote Post
SebastianG
post Mar 24 2004, 15:23
Post #4





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



so... how did you do it actually ?

bye,
Sebastian

This post has been edited by SebastianG: Mar 24 2004, 15:26
Go to the top of the page
+Quote Post
music_man_mpc
post Mar 24 2004, 15:32
Post #5





Group: Members (Donating)
Posts: 707
Joined: 20-July 03
From: Canada
Member No.: 7895



QUOTE (SebastianG @ Mar 24 2004, 06:23 AM)
so... how did you do it actually ?

bye,
Sebastian

See this thread.


--------------------
gentoo ~amd64 + layman | ncmpcpp/mpd | wavpack + vorbis + lame
Go to the top of the page
+Quote Post
SebastianG
post Mar 24 2004, 16:37
Post #6





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



QUOTE (music_man_mpc @ Mar 24 2004, 06:32 AM)
See this thread.


He never mentioned HOW he "tuned" it.
As long as he keeps his magic tuning a secret,
there's no reason for me to give a damn.

It's nothing personal. There are just enough ppl
out there thinking they know what they are doing.
So I keep being sceptic unless I get to know them
better.

bye,
Sebastian
Go to the top of the page
+Quote Post
tigre
post Mar 24 2004, 17:02
Post #7


Moderator


Group: Members
Posts: 1434
Joined: 26-November 02
Member No.: 3890



Here QuantumKnot said that he just forced lower -q levels to use the same stere mode as -q 10 does. I don't know if recent versions do something different, though.


--------------------
Let's suppose that rain washes out a picnic. Who is feeling negative? The rain? Or YOU? What's causing the negative feeling? The rain or your reaction? - Anthony De Mello
Go to the top of the page
+Quote Post
SebastianG
post Mar 24 2004, 17:15
Post #8





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



Ok, then. Sorry for the trouble.

If THAT's the only thing he did, I would not recommend
the usage of his binary. Anyway, everyone is free to do
whatever he likes wink.gif

bye,
Sebastian
Go to the top of the page
+Quote Post
QuantumKnot
post Mar 25 2004, 00:53
Post #9





Group: Developer
Posts: 1245
Joined: 16-December 02
From: Australia
Member No.: 4097



No, I didn't reuse the stereo modes of q 10 in this encoder. That was for a lossless coupled encoder for experimental purposes which is completely different to this one. Anyone who has actually encoded something will realise that the bitrates of let's say, q 4, of the HF reduced encoder are typically lower than what lossless coupled stereo would allow.

The source code for GT3b2 with HF reduction is out in the open and it is not my obligation to spend my time explaining every line of modified code. You seem to talk big yet you fail to actually put any effort into understanding what is actually done and what the listening results are telling. You make conclusions based on misleading and flawed observations yet never actually performed a listening test to verify the results.

I don't make any claims that I have solved or fixed the underlying problems in the Vorbis encoder. I only claim to have reduced the HF boost in certain transients and I ask people to use double blind listening tests to verify this. There are other situations (in very quiet passages) where there is a distinct HF boost in this encoder and I recognise that this wasn't solved. These points have been supported by numerous people who have performed proper ABC/HR tests, which is what this forum considers as legitimate evidence, not just looking at the frequency responses (which is a violation of TOS #8).

If you continue to pursue this direction of thinking that I am making fraudulent claims and all the work I have done is worthless, then that is not my problem. I put my trust in the double-blind listening tests my fellow forum members have voluntarily done. I never claim to have all the answers or that I have solved the problem. I only say I've reduced, not fix

- QK

This post has been edited by QuantumKnot: Mar 25 2004, 01:53
Go to the top of the page
+Quote Post
p0wder
post Mar 25 2004, 02:51
Post #10





Group: Members
Posts: 347
Joined: 22-July 02
From: USA
Member No.: 2721



QUOTE (SebastianG @ Mar 24 2004, 08:15 AM)
Ok, then. Sorry for the trouble.

If THAT's the only thing he did, I would not recommend
the usage of his binary. Anyway, everyone is free to do
whatever he likes wink.gif

bye,
Sebastian

If you can't look at the code and find the changes he made yourself, what gives you the right to not recommend it?
Go to the top of the page
+Quote Post
SebastianG
post Mar 25 2004, 17:20
Post #11





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



@QuantumKnot:

Don't get me wrong, it's nothing personal. We're both seem to be somewhat self-confident about our skills wink.gif You didn't want to spend some time to give some headwords for discussion able to describe the essentials of your tuning approach. And I'm too lazy to search for your source code version and to look for myself (that's why I asked you in my first reply)

"HF reduction" implies in my eyes that you've done something to the code to reduce the artificial boost effect. This I could not confirm for myself and obviously we've a different understanding in this field.
ABXing is a pretty good thing, but please be open-minded for other things as well. (It's ok to be sceptical first, though). You're saying it's a bad idea to generate noise and check for the average power of the quantized signal. I'm saying that the transients you're reffering to are just noise and the intensity depends on objectivly measurable energy.

I'm not saying your code is worthless or anything and I don't intend to do so until I'm aware of your tuning approach and have a good reason for this.
Because of tigre's reply I thought you're trying to fix the point-stereo thing by just avoiding it which sounds not that well for me.

Please understand my sceptism for ppl I don't know who try to tune an encoder. You may be as sceptical and think I'm a moron - which I probably am or not.

Maybe i'll find the time to look at your code and/or maybe you'll find the time to answer my first reply. If neither happens I guess there's no reason to further this discussion for me.

bye,
Sebastian

This post has been edited by SebastianG: Mar 25 2004, 17:27
Go to the top of the page
+Quote Post
tigre
post Mar 25 2004, 17:52
Post #12


Moderator


Group: Members
Posts: 1434
Joined: 26-November 02
Member No.: 3890



In my reply I said
QUOTE
I don't know if recent versions do something different, though.
So I don't feel guilty for this
QUOTE
Because of tigre's reply I thought you're trying to fix the point-stereo thing by just avoiding it which sounds not that well for me.

wink.gif

Besides this....

last time I checked, Vorbis was a codec for lossy audio compression. AFAIK such codecs are used for saving space and storing *audio* content for human listeners with *audibly* as good as possible reproduction of the original.

There's nothing wrong with using artificial test signals and measurements/graphs/numbers for testing, bugfixing etc., but in the end the audible result should tell what is a step in the right direction and what not. This is what I miss in your approach so far, SebastianG. Or have you actually listened to the test signals you've used for your measurements (-> ABX results wink.gif )?

edit: spelling of "lossy"

This post has been edited by tigre: Mar 25 2004, 17:58


--------------------
Let's suppose that rain washes out a picnic. Who is feeling negative? The rain? Or YOU? What's causing the negative feeling? The rain or your reaction? - Anthony De Mello
Go to the top of the page
+Quote Post
SebastianG
post Mar 25 2004, 20:36
Post #13





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



I don't feel guilty, too. I just said "IF this is the only thing he's changed, ...".
wink.gif

I completely understand your point, tigre and QK. This is my point: Do I have to rely on ABX tests when an unwanted effect is objectivly measurable ? Maybe we're still addressing different problems. I'm interested in preserving the energy level which I think is related to the increase of intensity of the mentioned transients.

There are 3 causes for this HF boost issue:
- low-SNR quantization with SNR maximizing thresholds (always valid)
- improper point-stereo code (valid for q<6)
- noise normalizing code (valid for q<6)

With the test i performed i only checked for the first cause. This explains the discrepancy of my test result to your ABX results if you did those tests for q<6 and QK has done something to reduce the effect of the point-stereo code.

bye,
Sebastian
Go to the top of the page
+Quote Post
QuantumKnot
post Mar 26 2004, 00:56
Post #14





Group: Developer
Posts: 1245
Joined: 16-December 02
From: Australia
Member No.: 4097



Just to clarify my position on graphs versus ABXing. I use frequency responses, graphs, spectrogram subtraction sometimes during my tunings to see what effect some modification has. However, in the end, I always do my own private listening tests to see if I have reduced something or not before I release it for external testing. And since my hearing and equipment is not that good, I often have to rely on other people like harashin, nyaochi, guruboolez, etc. to do ABX/ABC-HR tests to judge whether this encoder is an improvement or nor over the official encoder. Hence I believe, as tigre pointed out above, that empirical analysis such as spectrograms are only useful during tuning, bugfixing, etc. but ABX/ABC-HR tests should be used as the final judge of an encoder since that is ultimately what a perceptual encoder is used for.....compressing audio for our listening.

My impression is that you are judging my encoder based on looking at the frequency response of original/compressed white noise. In my opinion, that is not an accurate or fair way of judging any perceptual coder. Sure, looking at these graphs can reveal the underlying behaviour but when it comes to perceptual quality, nothing can replace ABX/ABC-HR testing.

I don't claim I know everything or a lot about Vorbis internals. I am not a Vorbis developer by trade and I admit I don't know half of what people like Garf, nyaochi, aoyumi, etc know about Vorbis internals. I stress time and time again in the other threads that I am learning as I go. I have 30% confidence that my tunings will actually work until I get external testers who very graciously perform listening tests with their expertise and equipment. Only till then am I moderately confident, based on their comments. The HF reduction code in this version of GT3b2 came from my own tuned version, QKTune beta 3.2. Since I got positive comments from the listening tests people did that there was a degree of HF hiss/boost reduction in certain transients, compared with the original coder, then I applied the patch to GT3b2, as a proof of concept.
Go to the top of the page
+Quote Post
SebastianG
post Mar 26 2004, 10:49
Post #15





Group: Developer
Posts: 1318
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



QUOTE (QuantumKnot @ Mar 25 2004, 03:56 PM)
... Hence I believe, as tigre pointed out above, that empirical analysis such as spectrograms are only useful during tuning, bugfixing, etc. but ABX/ABC-HR tests should be used as the final judge of an encoder since that is ultimately what a perceptual encoder is used for.....compressing audio for our listening.

Totally agree.

QUOTE
My impression is that you are judging my encoder based on looking at the frequency response of original/compressed white noise.  In my opinion, that is not an accurate or fair way of judging any perceptual coder.

Agree, it isn't. I'm using it *only* for judging concerning the energy preservation issue. I don't complain about a lowpass-filtered signal or whatever ...

I really do appreciate your efforts. Though, it would have been cool, if we had discussed the tuning approach itself.

bye,
Sebastian
Go to the top of the page
+Quote Post
Biont
post Apr 1 2004, 23:36
Post #16





Group: Members
Posts: 53
Joined: 6-May 03
Member No.: 6439



QK! Your GT3b2 with HF reduction lacks multilanguage support. My Russian files could not be interpreted. Could you do something with it? Plus remove the line, which says "Experimental codec, do not use for archiving"?
Go to the top of the page
+Quote Post
QuantumKnot
post Apr 2 2004, 00:34
Post #17





Group: Developer
Posts: 1245
Joined: 16-December 02
From: Australia
Member No.: 4097



QUOTE (Biont @ Apr 2 2004, 08:36 AM)
QK! Your GT3b2 with HF reduction lacks multilanguage support. My Russian files could not be interpreted. Could you do something with it? Plus remove the line, which says "Experimental codec, do not use for archiving"?

Try John33's compile of OggDropXPd in the other GT3b2 with HF reduction thread. I think that has multilanguage support (I might be wrong).
Go to the top of the page
+Quote Post

Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 22nd November 2014 - 22:56