Lame 3.99 320 CBR Killer Sample?, Split From Topic ID: 100182 (TOS #5) Options

[Propheticus]

If it's really about quality I'd go with a lossless format. Doesn't really matter which, but lets say FLAC for compatibility reasons. I have been able to ABX FLAC against MP3 (LAME) CBR 320kbps V0 multiple times now.
The MP3 was a LAME convert I made with Foobar2000. To be sure the levels matched I rescanned the mp3 with Replaygain. But unsurprisingly the adjustment stayed the same. The difference is most clear to me in the highs (cymbals/harsh transients/sibilants in text) which sound harsher to me in mp3's. I checked this afterwards, the ABX was blind ofc.

info FLAC file:
Bitrate : 885 kbps
Track gain : -6.86 dB
Track peak : 0.966888

info MP3 file:
Bitrate : 320 kbps
Codec profile : MP3 CBR
Tool : LAME3.99r
Track gain : -6.86 dB
Track peak : 1.010323

CODE

foo_abx 1.3.4 report
foobar2000 v1.2.4
2013/04/02 21:09:26

File A: Z:\...\Paul Simon - Graceland - (2012 ) - (25th Anniversary Edition)[FLAC]\06 - You Can Call Me Al.flac
File B: Z:\...\006. Paul Simon - You Can Call Me Al -mp3-320.mp3

21:09:26 : Test started.
21:11:16 : 01/01 50.0%
21:12:05 : 02/02 25.0%
21:12:47 : 03/03 12.5%
21:13:31 : 04/04 6.3%
21:13:51 : 05/05 3.1%
21:14:27 : 06/06 1.6%
21:14:45 : 07/07 0.8%
21:15:11 : 08/08 0.4%
21:15:16 : Test finished.

----------
Total: 8/8 (0.4%)

btw I stopped at step 8 because p=0.04 is statistically significant (when taking α=0.05) and I had the feeling I could continue to discern the two. It took some effort but sometimes I heard which was which within 20 seconds.

This post has been edited by Propheticus: Apr 2 2013, 20:35

[pdq]

It would appear that your mp3 file is clipping. Perhaps that is the difference that you hear. Or am I misinterpreting the information that you gave?

[Propheticus]

I can find no evidence of clipping when running a "find clipping" analysis in Audacity. Also zooming at some peaks myself showed no clipping (other than near clipping parts that are near clipping on the original as well).


Above is the MP3.
Below the original FLAC.

It didn't sound like clipping either, a lot more subtle.

This post has been edited by Propheticus: Apr 2 2013, 21:14

[Dynamic]

Good to be on the look out to avoid clipping, but I don't think 1.01 peak in the whole file is enough clipping to be audible (but I can't ABX that from here!).

In any case, using foo_abx, you can use ReplayGain to match volume, which will remove clipping. It's quite likely it only clipped for one sample at that value.

I've only been told of one occasion some years ago where MP3 decoder-generated clipping was audible to somebody and successfully ABXed, and more normal peak levels are 1.1 to 1.3 (i.e. in the region of 20 to 25 dB more clipping distortion to a first approximation), and even those are hardly ever audible because of the types of sound that usually accompany them. Most LAME settings were tuned for transparency at full volume, despite decode-side clipping and despite the ongoing loudness war.

Assuming there's no odd 'effects' in the signal chain this could well be a genuine problem sample (possibly only a subtle one), which I applaud Propheticus for testing and bringing to our attention. It seems likely to be worth a look at replicating the result.

It would be interesting to test a <30 second sample of the affected part of that edition of Graceland (you could export a short 'selection' losslessly from Audacity and uploads it to the Uploads forum and we could attempt to replicate your successful ABX, Propheticus, especially if you could narrow down the timings or descriptions of the problem you most clearly notice).

I've only got the Spotify version of that release available (Vorbis encoded), but it certainly contains some sharp transients and possibly some simultaneous tones. This type of problem is often solved by halb27's extension of LAME with its -Vn+ settings which allow the momentary bitrate to greatly exceed 320 kbps in short blocks by using maximum available bit reservoir. I suspect -V2+ or -V0+ could well do the trick where 320 kbps CBR currently fails, as it has on a number of other samples.

Anyway, testing of that sample is probably best saved for another thread.

[2Bdecided]

I can find no evidence of clipping when running a "find clipping" analysis in Audacity. Also zooming at some peaks myself showed no clipping (other than near clipping parts that are near clipping on the original as well).

I'd find a better mastering (if one exists) before worrying about mp3-encoding it.

That digital clipping should not be on a recording from 1986!

Google found...
http://forums.stevehoffman.tv/threads/paul...11-sony.271333/
...which suggests the original mastering is the best (and the one you have is probably the worst).

Cheers,
David.

This post has been edited by 2Bdecided: Apr 3 2013, 15:09

[Propheticus]

Be that as it may... I was still able to tell the difference between flac and mp3. Does the source really matter then? Sources will be sub-optimal at times.

I found the 1986 master... same shape of near-clipping parts. Again my samples did not clip (I analysed in Audacity). The mp3 had 1 clip warning at a point far beyond the clip I was listening.

This post has been edited by Propheticus: Apr 3 2013, 16:07

[2Bdecided]

Be that as it may... I was still able to tell the difference between flac and mp3. Does the source really matter then? Sources will be sub-optimal at times.

It's relevant and interesting if 320kbps mp3 can be ABXed. I wasn't denying that.

It's far less "important" if it's due to clipping (or, for example, transcoding from a lossy source) - these are two special cases (there may be more) where we expect it to fail. Given that it's neither of these, then either you're exceptional, of you've found a problem sample. Yes, that matters.


My point was different: if you're the sort of person who is going to ABX 320kbps mp3 and decide to use lossless instead (and maybe even someone else is reading and they're not this kind of person, but they do care enough about audio quality to read HA), the first thing to do is to find the best available version to work from. That can make far more of a difference to the audio quality than mp3 vs FLAC. IMO. YMMV.

Cheers,
David.

This post has been edited by 2Bdecided: Apr 3 2013, 16:09

[Propheticus]

As per request I've uploaded a small sample of the FLAC file I used (uploaded sample). This sample is the bit I set the starting point at when I was ABX'ing.

You're right. Source material decides the SQ over all other aspects. Garbage in = Garbage out.

This post has been edited by Propheticus: Apr 3 2013, 16:22

[db1989]

I have been able to ABX FLAC against MP3 (LAME) CBR 320kbps V0 multiple times now.

Which is it? -b320, -V0, or the occasionally mentioned hybrid -V0 -b320?

[Propheticus]

Hhm somehow showed v0 somewhere, twas CBR not V0. Sorry, my bad.

CODE

Audio
Format : MPEG Audio
Format version : Version 1
Format profile : Layer 3
Mode : Joint stereo
Duration : 4mn 40s
Bit rate mode : Constant
Bit rate : 320 Kbps
Channel(s) : 2 channels
Sampling rate : 44.1 KHz
Compression mode : Lossy
Replay gain : -6.86 dB
Replay gain peak : 1.010323
Stream size : 10.7 MiB (100%)
Writing library : LAME3.99r
Encoding settings : -m j -V 4 -q 3 -lowpass 20.5

I used Foobar2000 and chose the 'highest' mp3 preset. No other processing, only replaygain scan afterwards.

This post has been edited by Propheticus: Apr 3 2013, 17:05

[db1989]

Thanks for the confirmation and additional info.

[eahm]

Propheticus, could you please test AAC and Opus? I'd love to know how low you can go.

It seems you are more trained to hear artifacts.

Thanks.

[Dynamic]

I took a listen and used fb2k's Converter to convert using lame 3.99r at the simple 320 kbps slider setting. I removed all ReplayGain info and played at full volume, using DirectSound output with Best quality resampling and no effects in Windows 7's audio pipeline over Philips SHE2550 earbuds.

I could not pick out anything around the snare and hi-hat around seconds 4-6, so I tried the last few seconds. Thought I was getting somewhere with the sharpness of the claps or rim hits on the beats near the end, but ended up 5/10. Went back to the 4-6 second area and thought maybe the onset of the brass note after the hi hat was possible to discern, but couldn't ABX that either despite attempting to pause, regain concentration and pick my focus. It seems that what I thought might be subtle differences, were probably my imagination.

I should note that I'm not especially good with transients unless they contain tonal elements that lack pitch precision, so while my attempt does not confirm your ABX result, nor does it offer evidence against it. I'm not a very well-trained artifact spotter.

It's also slightly possible that the timing of the cut can affect where in each transform window certain sounds occur, so my MP3 might differ significantly from yours. I've attached mine in the same Upload thread.

I might give it another go tomorrow when I'm fresh in case I can spot something, but if anyone else has results to report, I'm sure they'd help.

[Propheticus]

What kind of settings would be fair/comparable to MP3 CBR 320kbps? I'm tot really familiar with those formats.

AAC(nero) VBR q=0.79 for instance has a ~320kbps bit rate, would this be appropriate?

Just tried AAC LC CBR 320kbps... Was not able to tell a difference at all.

This post has been edited by Propheticus: Apr 3 2013, 19:58

[Dynamic]

Encoding Settings : -m j -V 4 -q 3 -lowpass 20.5

I don't know which tool gave that info, but is the -V 4 part normal for 320kbps CBR, anyone? Feel free to download my version and compare its Encoding Settings with the same tool.

I'm aware that recent LAME is using a lot of the VBR psychoacoustics in 320kbps CBR mode to better use sfb21 etc, but I'd have been surprised it wasn't -V 0.

[Propheticus]

This info was derived from the program 'MediaInfo'.
Your sample shows the same info btw.

This post has been edited by Propheticus: Apr 3 2013, 20:10

[eahm]

Just tried AAC LC CBR 320kbps... Was not able to tell a difference at all.

Pleas, that's way too much and not the encoder I would test. If I may, download TAudioConverter, set the encoder to AAC/qaac (Apple) and test the following TVBR settings:

36 = ~95 kbps
63 = ~135 kbps
73 = ~150 kbps
82 = ~165 kbps
91 = ~195 kbps
100 = ~225 kbps
109 = ~255 kbps
118 = ~285 kbps
127 = ~320 kbps

Try from the lowest and stop when you don't ear any difference. Thanks.

edit:
I can convert the sample for you if you'd like me to.

This post has been edited by eahm: Apr 3 2013, 20:16

[db1989]

I don't know which tool gave that info, but is the -V 4 part normal for 320kbps CBR, anyone?

AFAIR, this has been brought up before, and the -V setting is not relevant for CBR; possibly the program even just uses other methods to ‘guess’ the value of the parameter when it has not been directly applied.

I'm aware that recent LAME is using a lot of the VBR psychoacoustics in 320kbps CBR mode to better use sfb21 etc, but I'd have been surprised it wasn't -V 0.

I don’t know whether or not some VBR routines are being incorporated into CBR modes, or to what degree. In any case, 320 kbps CBR as encoded by -b320 still is not the same as -V0 -b320. The latter engages the full VBR engine and might provide benefits compared to the former in certain contexts, although it is still unnecessary to use anything other than the normal/recommended configuration in the majority of cases.

This post has been edited by db1989: Apr 3 2013, 20:18

[Propheticus]

When I'll come to test this again (some other day, had my share of the "you can call me al" sample today ) I'll start with the 127 setting. If this ABX is a positive, this saves me from doing 8 tedious ABX runs.

[halb27]

In case the issue is due to pre-echo I suggest to use -V0 instead of CBR320.
If you like to try my functional extension 3.100i I suggest you use something between -V2+ and -V0+ as was suggested before by Dynamic.

ADDED:
I tried to ABX 3.99.5 -b 320 but couldn't (doesn't say much in case it's a pre-echo issue).
I also analyzed the track with the --frameAnalyzer option of 3.100i: most of the time granule sequences of type 'short' contain just one or two granules, so 3.100i -V0+ has a high potential of improving over -V0 in case -V0 doesn't do the job already.

So I welcome very much if you could test -V0 and - in case there's still something wrong - 3.100i -V0+.

This post has been edited by halb27: Apr 3 2013, 21:30