LAME MP3: Strange "Spectre" - LAME change the spectral messa Options

[Antigen]

Hi to all,

I have compared the audio spectre of a LAME MP3 with a FLAC obtained from an original CD.

The result is this for MP3 LAME:



Uploaded with ImageShack.us


and this is the result obtained from a FLAC (lossless)



Uploaded with ImageShack.us



As can you see, the sound message into the MP3 has changed not only into the frequency, but into the "compression" in dB.

I see that FLAC has a compression of - 54 dB and Lame of -90 dB, a lost of Dynamic?

I have tried the same test with AAC and the FLAC don't change the "Dynamic" in terms of dB.

Why LAME change the dB of the sound message?

This post has been edited by Antigen: Jun 26 2012, 07:51

[saratoga]

I think you are misreading the scales on those images.

[dhromed]

You probably have a different window function selected in the screenshots. The graphs are pretty much the same.

[db1989]

Even if your graphs were matched for scale:

Graphs, non-blind listening tests, waveform difference comparisons, and so on, are not acceptable means of providing support.

Audio is consumed by ears, not eyes. Visual differences in perceptually encoded audio are to be expected and are not to be spuriously correlated with supposed differences in quality. In fact, it would be the mark of a highly effective lossy encoder if it could change some visual representation ‘significantly’ without being audibly different at all.

TL;DR: We do not care what graphs look like. We care what waveforms sound like. Please bear this in mind.

[lvqcl]

The graphs are more or less the same for 0...20kHz range.

[dhromed]

Visual differences in perceptually encoded audio are [...] not to be spuriously correlated with supposed differences in quality.

I don't read any claims of that nature in Antigens post. He/she is merey asking why these graphs look so different, which is an entirely valid question and, I think, easily explained with misreading the axes and/or having accidentally selected a different window function in Audacity.

[db1989]

Perhaps I am being reflexively cautious, and I’m not accusing anyone of anything, but I’d like to preclude any unrealistic assumptions being taken from graphs. Short of demonstrating that the sound has indeed changed and that lossy encoding alters the balance of different frequency ranges, what substantive use do they have?

You’ll also excuse me if I want to drive this point home in case Antigen has ‘heard’ somewhere on the internet that visual representations are a valid method of evaluation. I mean no offence there, but a sizable history of too readily taking claims on faith seems to make it important to emphasise that this isn’t the case.

[Antigen]

Perhaps I am being reflexively cautious, and I’m not accusing anyone of anything, but I’d like to preclude any unrealistic assumptions being taken from graphs. Short of demonstrating that the sound has indeed changed and that lossy encoding alters the balance of different frequency ranges, what substantive use do they have?

You’ll also excuse me if I want to drive this point home in case Antigen has ‘heard’ somewhere on the internet that visual representations are a valid method of evaluation. I mean no offence there, but a sizable history of too readily taking claims on faith seems to make it important to emphasise that this isn’t the case.

I don't have heard this from anyone.

I have simply make this procedure:

1) convert a CD into MP3 with XLD (CBR 320)

2) convert a CD into FLAC

3) convert a CD into AAC

Open the first track with AUDACITY and make the graph that I posted.

Open the second track with AUDACITY and make the graph that I posted.


The graph talk, not me.

Someone can explain the motivation of the variation in dB?

I don't talk about the quality of the sound!

I want to know only because LAME change the dynamic of the graph

This post has been edited by Antigen: Jun 26 2012, 13:14

[lvqcl]

I want to know only because LAME change the dynamic of the graph

It DOESN'T. The dynamic of the graph? What is it?

This post has been edited by lvqcl: Jun 26 2012, 13:59

[db1989]

I don't have heard this from anyone.

I have simply make this procedure:

1) convert a CD into MP3 with XLD (CBR 320)

2) convert a CD into FLAC

3) convert a CD into AAC

Open the first track with AUDACITY and make the graph that I posted.

Open the second track with AUDACITY and make the graph that I posted.

Why?

The graph talk, not me.

The graphs are not using the same vertical/y-axis scale, as has been said numerous times already. So, in terms of anything approaching a valid comparison between the two, they don’t “talk” about anything.

Someone can explain the motivation of the variation in dB?

Differing y-axes.

I don't talk about the quality of the sound!

Good.

I want to know only because LAME change the dynamic of the graph

Differing y-axes.

Even if equating the vertical scales and re-comparing did show anything substantial, which dhromed has said is not the case, it wouldn’t matter. Hydrogenaudio does not accept visual ‘evidence’ of sound quality. Since you have said that talk of quality is not your intention, I must again reiterate (1) my question of why you did this at all and (2) the fact that differences that don’t affect quality — (choose one:) are irrelevant to / are a large part of the entire purpose of — lossy compression.

[Antigen]

I have done another test:

- a LAME VBR -V0

- a LAME CBR 320

Why with -V 0 if I analyze the fruency I see that are present frequency at 21.000 Hz and at 320 CBR there is cutoff at 20.000?

Possible that VBR can make a better audio representation?

[benski]

I have done another test:

- a LAME VBR -V0

- a LAME CBR 320

Why with -V 0 if I analyze the fruency I see that are present frequency at 21.000 Hz and at 320 CBR there is cutoff at 20.000?

Possible that VBR can make a better audio representation?

VBR -V0 is likely better than CBR 320 (objectively and subjectively) due to different bit allocation methods. This has been discussed before on hydrogenaudio. However, except for a few "problem" samples (such as halb27's), it is impossible to verify this claim as both of these settings produce a quality much higher than expert listeners can discern.

Note that trying to analyze quality by frequency representation isn't correct. This is why the old Blade encoder sounded bad. There's a lot more to the story (such as phase response and masking effects). Automated simulated-listening-tests from audio analysis has been discussed here before (It has a name, but I can't remember it for the life of me). It's no substitute for human ears, and probably only good for an initial pass on codec tuning during early development.

Also, the difference in your two graphs is just a quirk in your program that tries to auto-scale the axes. As an example, these two graphs show identical information, but the Y-axis has been selected differently.





This post has been edited by benski: Jun 26 2012, 21:12

[db1989]

Note that trying to analyze quality by frequency representation isn't correct.

Quoted for (re-)emphasis. Please try searching if you want to read any of the countless past discussions about lowpass filtering in MP3, often with specific reference to LAME and its differing modes. Once again, I imagine that most questions you could ask about the topic have already been answered elsewhere.

[greynol]

Looks like there's been some progress from "will you tell me what are good settings?" to "I've taken some initiative in order to determine if there are quality differences."

Unfortunately the progress is in the wrong direction.

http://www.hydrogenaudio.org/forums/index....showtopic=16295

BTW, here's a version of the mp3 graph that has been scaled so that apples can once again be compared to apples rather than to fruit loops:


This post has been edited by greynol: Jun 27 2012, 14:20

[halb27]

It's a good example of the fact that
- our judgement about things is controlled more by emotion triggered by the bird's view than by reasoning about substantial details
- the big picture of the graph as a whole controls our emotion ignoring scale.

In another context inappropriate scaling in graphical presentations is often used willingly to mislead the judgement of the readers in the sense of the author as demonstrated by benski's post.
(BTW I often felt like this when seeing a 'zoomed' version of a listening test result).

This post has been edited by halb27: Jun 27 2012, 07:41

[greynol]

"The result was a tie."
"Yes, I know. So which one won?"

...not unlike
"They are both transparent."
"OK, so which one sounds better?"

IOW, "But these amps go to eleven."

[Fedot L]

LAME change the spectral messa

Nothing of the kind!
Very easy to verify it looking, for example, at the levels of two points of the spectrum on both graphs:
5800 Hz: minus 38,5 dB both graphs;
19000 Hz: minus 43 dB, both graphs,
etc etc etc…