IPB

Welcome Guest ( Log In | Register )

 
Reply to this topicStart new topic
lossyWAV vs MP3
chrizoo
post Sep 13 2011, 07:40
Post #1





Group: Members
Posts: 345
Joined: 25-March 08
Member No.: 52274



Hi. I very much like the theoretical concept of lossyWAV. But being skeptical by nature, I always like to try for myself and verify.

This is my test set-up:
file1: the source wav file
file2: file1 processed with lossyWAV --extreme
file3: file1 compressed with LAME mp3 320CBR and decoded back to wav
file4: file1-file2
file5: file1-file3

Results:
Regarding the waveform, I generally noticed,
that file4 is always "centered" at the zero line (while file5 strays from it), i.e. positive and negative half of the wave are quite symmetrical
but
that file4 generally has considerably higher amplitudes than file5.
Hereafter is a screenshot sample to illustrate that.

sample from file1 which includes the part shown in the screenshot:
http://www.hydrogenaudio.org/forums/index....showtopic=90776


Now the question is:
While mp3 (of any bitrate) is generally not recommended to transcode to another lossy format, lossyWAV --extreme is:

QUOTE (http://wiki.hydrogenaudio.org/index.php?title=LossyWAV)
--quality extreme: (-q E or -q 7.5) Higher quality preset, disc space-saving alternative to lossless archiving for large audio collections, considered to be suitable for transcoding to other lossy codecs;

So the lossyWAV's loss should without doubt be considerably less than the loss/distortions introduced by mp3.
The comparison doesn't exactly suggest that (if anything, it would suggest the contrary).

On the one hand waveform comparison allegedly cannot reliably inform about quality loss (according to HA),
but on the other hand I am asking myself HOW it can be possible, that file4 COULD even somehow be closer to the original file (than file5) when we see that it is in fact more different?



This post has been edited by chrizoo: Sep 13 2011, 07:45
Go to the top of the page
+Quote Post
Northpack
post Sep 13 2011, 08:21
Post #2





Group: Members
Posts: 455
Joined: 16-December 01
Member No.: 664



QUOTE (chrizoo @ Sep 13 2011, 07:40) *
Now the question is:

So the lossyWAV's loss should without doubt be considerably less than the loss/distortions introduced by mp3.
The comparison doesn't exactly suggest that (if anything, it would suggest the contrary).

Why? A difference graph doesn't say anything about the audibility of the distortions it reflects.

QUOTE
On the one hand waveform comparison allegedly cannot reliably inform about quality loss (according to HA),
but on the other hand I am asking myself HOW it can be possible, that file4 COULD even somehow be closer to the original file (than file5) when we see that it is in fact more different?

The answer is simple: because what you see is not what you hear. If you want a more evident example: take a waveform sampled at 192khz (containing frequency content up to 96khz) and perform low pass at 24khz. You've thrown away 3/4 of all data. And yet I bet you wouldn't hear the slightest difference.

lossyWAV has an approach very different from MP3, it is a "variable bitrate preprocessor" which essentially estimates how many bits can be chopped off (zeroed) for a given block of audio at a given quality level. Those chopped-off bits will always directly show up in the difference graph, while the changes MP3 compression introduce to the audio don't necessarily have a big effect on the difference graph and yet can be very well audible.

Comparing those graphs is like comparing apples and oranges and saying: the oranges are bigger so they must be more healthy wink.gif

This post has been edited by Northpack: Sep 13 2011, 08:38
Go to the top of the page
+Quote Post
chrizoo
post Sep 13 2011, 08:36
Post #3





Group: Members
Posts: 345
Joined: 25-March 08
Member No.: 52274



QUOTE (Northpack @ Sep 13 2011, 08:21) *
A difference graph doesn't say anything about the audibility of the distortions it reflects.

That is self-evident! Something visual can never say whether something is audible.
But the difference graph tells us how much difference there is from the original, no?

QUOTE
I bet you wouldn't hear the slightest difference.

The point of this thread isn't whether the mp3 differences and the lossywav differences are beyond the hearing threshold,
but
which file is closer/more faithful to the original (i.e. has less difference).

This post has been edited by chrizoo: Sep 13 2011, 08:37
Go to the top of the page
+Quote Post
greynol
post Sep 13 2011, 08:38
Post #4





Group: Super Moderator
Posts: 10256
Joined: 1-April 04
From: San Francisco
Member No.: 13167



QUOTE (chrizoo @ Sep 13 2011, 00:36) *
closer/more faithful to the original (i.e. has less difference)

"Less difference" is irrelevant when it comes to lossy compression.

This post has been edited by greynol: Sep 13 2011, 08:41


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
chrizoo
post Sep 13 2011, 08:42
Post #5





Group: Members
Posts: 345
Joined: 25-March 08
Member No.: 52274



QUOTE (greynol @ Sep 13 2011, 07:38) *
"Less difference" is irrelevant when it comes to lossy compression.

Well, with lossless compression, there is no difference at all, so the distinction "when it comes to lossy compression" becomes irrelevant and can be stripped,
thus your post says:

QUOTE
"less difference" is irrelevant

I don't understand why.

This post has been edited by chrizoo: Sep 13 2011, 08:43
Go to the top of the page
+Quote Post
greynol
post Sep 13 2011, 08:46
Post #6





Group: Super Moderator
Posts: 10256
Joined: 1-April 04
From: San Francisco
Member No.: 13167



You were already told the answer.

Faithful reproduction does not directly correlate to differences in amplitude with perceptual encoding.

As far as this forum is concerned, faithful reproduction of a perceptual encoding is to be judged through double-blind testing.

This post has been edited by greynol: Sep 13 2011, 08:57


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
Northpack
post Sep 13 2011, 08:51
Post #7





Group: Members
Posts: 455
Joined: 16-December 01
Member No.: 664



QUOTE (chrizoo @ Sep 13 2011, 07:36) *
QUOTE
I bet you wouldn't hear the slightest difference.

The point of this thread isn't whether the mp3 differences and the lossywav differences are beyond the hearing threshold,
but which file is closer/more faithful to the original (i.e. has less difference).

Your question was:
QUOTE
I am asking myself HOW it can be possible, that file4 COULD even somehow be closer to the original file (than file5) when we see that it is in fact more different?

I've answered that question. The same is true for transcoding from one lossy format to another. The reason why lossyWAV is considered to be more adequate for transcoding is because it has a very simple psychoacoustical model that is unlikely to interfere with the more sounded-up psy-models of other lossy encoders. There's some empirical evidence for that claim, but as long as there hasn't been any properly conducted listening test it's just an assumption.
Go to the top of the page
+Quote Post
chrizoo
post Sep 13 2011, 09:21
Post #8





Group: Members
Posts: 345
Joined: 25-March 08
Member No.: 52274



Look, of course none of you is forced to explain anything to me or to talk about the waveform at all. But simply saying "only listening to differences is accepted here, but looking at the waveform is not", does not make me (or anyone else who is looking for answers) understand anything. I've been here long enough to know that the consensus here is that only listening matters (But WHY that is so is a different matter). However, I find the theoretical background interesting too (plus I don't hear so well).

It's not like the waveform doesn't depict exactly the acoustic signal we are dealing with here. And if in the difference graph, the amplitude is higher, it means, there is more difference from the original audio. Correct me if I am wrong, but I think that is a fact.

The answers included explanations of how lossyWAV works - which I think I broadly know already - but this is irrelevant (because not intrinsic) to explaining why difference graphs in general don't matter. And this point just is not understandable to me (so far).

This post has been edited by chrizoo: Sep 13 2011, 09:27
Go to the top of the page
+Quote Post
Northpack
post Sep 13 2011, 09:51
Post #9





Group: Members
Posts: 455
Joined: 16-December 01
Member No.: 664



It seems to me that you don't want to understand what greynol and I explained. I suggest that you open your mind and read my first post in this topic again, it's all said in there. I haven't declared any dogmns, it's just good old plain logic.

Of course there is some relation between the perceived quality and the difference graph, but it's no direct relation, it's heavily complex, nonlinear and dependent from many unknown variables.

If you still don't understand I propose that you do a very simple experiment. Take a waveform and invert it in audacity. Let's say the inverted waveform is the encoder's output. It sounds exactly identical to the original. Now go and calculate the difference function.

This post has been edited by Northpack: Sep 13 2011, 10:00
Go to the top of the page
+Quote Post
greynol
post Sep 13 2011, 09:51
Post #10





Group: Super Moderator
Posts: 10256
Joined: 1-April 04
From: San Francisco
Member No.: 13167



QUOTE (chrizoo @ Sep 13 2011, 01:21) *
And if in the difference graph, the amplitude is higher, it means, there is more difference from the original audio.

I realize this doesn't help you understand why or how, but when it comes to perceptual encoding, audibility and efficiency are king. Differences from original audio are measured through tests of audibility, not through differences in amplitude. With this in mind, the quoted claim is not necessarily true. If it turns out that the lossywav version of your example is transparent while the mp3 is not then your recent discovery can only serve to reinforce this fact.

This post has been edited by greynol: Sep 13 2011, 10:03


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
dhromed
post Sep 13 2011, 11:14
Post #11





Group: Members
Posts: 1337
Joined: 16-February 08
From: NL
Member No.: 51347



QUOTE (chrizoo @ Sep 13 2011, 10:21) *
simply saying "only listening to differences is accepted here, but looking at the waveform is not", does not make me understand anything


There exist artifacts that are easily seen but not heard. There exist artifacts that are easily heard but hard to see.
Therefore looking at a waveform conveys very little useful information about how it sounds.

A waveform can give a likely idea how loud it is, and how the (de)crescendos are structured. That is basically all. (incidentally the reason why people enjoy foobar's waveform seekbar component smile.gif )

QUOTE (chrizoo @ Sep 13 2011, 10:21) *
It's not like the waveform doesn't depict exactly the acoustic signal we are dealing with here.


The waveform is a graph of intended air pressure changes. It is an abstraction. No visual representation can exactly depict the acoustic signal.

QUOTE (chrizoo @ Sep 13 2011, 10:21) *
And if in the difference graph, the amplitude is higher, it means, there is more difference from the original audio. Correct me if I am wrong, but I think that is a fact.


Technical fact, yes, but also meaningless as to how it sounds in the end. See Northpack's waveform inversion for a technical difference of 100% with 0% change in audibility.

This post has been edited by dhromed: Sep 13 2011, 11:14
Go to the top of the page
+Quote Post
alexeysp
post Sep 13 2011, 11:17
Post #12





Group: Members
Posts: 141
Joined: 3-April 09
Member No.: 68627



QUOTE (chrizoo @ Sep 13 2011, 11:21) *
It's not like the waveform doesn't depict exactly the acoustic signal we are dealing with here. And if in the difference graph, the amplitude is higher, it means, there is more difference from the original audio. Correct me if I am wrong, but I think that is a fact.


You seem to overlook one very important thing: the human auditory system does not act as a waveform analyzer, but rather as a spectrum analyzer. Moreover, this spectrum analyzer is non-linear, since it demonstrates considerable masking effects. That's why waveform analysis (and waveform difference analysis) is not very useful when it comes to perceptual encoding. While waveform does depict the signal, it doesn't depict the features of the signal that we actually perceive.
Go to the top of the page
+Quote Post
saratoga
post Sep 13 2011, 19:01
Post #13





Group: Members
Posts: 5118
Joined: 2-September 02
Member No.: 3264



QUOTE (chrizoo @ Sep 13 2011, 04:21) *
I've been here long enough to know that the consensus here is that only listening matters (But WHY that is so is a different matter). However, I find the theoretical background interesting too (plus I don't hear so well).


The ears are basically a nonlinear spectrometer, reading out spectral differences in waveforms, while the eyes are spatial domain instruments, reading out spatial variations with little spectral resolution. If you try to interpret audio data in a domain other then the one its measured in, the results are likely to be weird if not completely meaningless because not all differences in one domain will map into changes in measurements in the other.
Go to the top of the page
+Quote Post
mjb2006
post Sep 13 2011, 21:19
Post #14





Group: Members
Posts: 848
Joined: 12-May 06
From: Colorado, USA
Member No.: 30694



QUOTE (dhromed @ Sep 13 2011, 04:14) *
There exist artifacts that are easily seen but not heard. There exist artifacts that are easily heard but hard to see. Therefore looking at a waveform conveys very little useful information about how it sounds.

I would like to underscore and boldface this point. Another way to put it is that the difference between two waveforms can visually look very big but sound very small (nil in the example of the inverted waveform), and vice-versa.
Go to the top of the page
+Quote Post
2Bdecided
post Sep 14 2011, 10:14
Post #15


ReplayGain developer


Group: Developer
Posts: 5279
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



What's the vertical scale on your graph chrizoo? It's not LSBs/bits, and it's not wrt full scale.

Masked noise ~25dB below the peak signal level can be inaudible.
Unmasked (e.g. white) noise ~48dB below the peak signal level is easily audible in many cases, and even ~60dB below is audible sometimes.

Of course the former looks larger than the latter if you simply plot the noise (original minus coded = difference) waveform. The latter is often invisible unless you zoom in.

Any process that adds large quantities of ultrasonic noise (e.g. DSD) can yield a difference plot that's larger than original signal, while sounding identical.

Cheers,
David.
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: 26th November 2014 - 01:16