Skip to main content

Topic: Possible Poor Performance Of Alt-presets (Read 12602 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.
  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
I've just doing some ABX tests with the sample trumpets1.wav, which you can find at http://www.pcabx.com/product/reference/trumpets1.wav

It's a trumpet sound, with no pre-echo issues, just added harshness on some alt-presets.

My results:

--alt-preset standard: 6/6 p= 1.6%, quite easy
--alt-preset standard -Z : 7/7 p= 0.8%, quite easy
--alt-preset extreme : 7/7 p= 0.8%, quite easy
--alt-preset extreme -Z : 12/14  p= 0.6%, seems more difficult, or I'm more tired
--alt-preset insane: 3 trials, total 20/32 p=14%, quite difficult, I think I can ABX it on optimum conditions, but I can't prove it right now  . Edit: not so sure I can ABX it...





So, seems that the VBR alt-presets have more room for tuning?

  • Dibrom
  • [*][*][*][*][*]
  • Administrator
Possible Poor Performance Of Alt-presets
Reply #1
There are a few type of samples where there are still some problems with --alt-preset standard, yes.  They are not extremely common though.

As for the cause of the problem, that basically resides at the core of nspsytune, noise shaping type 2, and the different quantization comparison methods.  There is a whole lot of room for improvement in the core psyacoustics of LAME.  Many of these things were going to be addressed in the now defunct nspsytune2 (Naoki Shibata has virtually disappeared from the scene for what I understand to be some personal reasons...).  I'm not sure if they'll ever be fixed now.  They aren't really things that I have enough knowledge to fix by myself either.  The best I can do is try to bypass as many of the existing flaws as possible, and that's pretty much what the --alt-presets already.

In summation, I don't think the --alt-presets are going to get much better without the core psymodel of LAME being massively overhauled first 

MP3 really isn't the best choice these days if you want the best quality in all (or nearly all) cases.  Most of the other newer formats have much better performance and a higher degree of reliability for delivering a set level of quality.

However, I do appreciate your input

  • fewtch
  • [*][*][*][*][*]
Possible Poor Performance Of Alt-presets
Reply #2
Dibrom,

Are you basically stating here that due to circumstances, it's unlikely the --alt-presets can/will ever be further improved (significantly) in terms of quality?

As for the other formats, unless/until they have some significant hardware support they are nearly useless, IMHO (who wants to sit in a hard desk chair in front of a computer, listening to music?).  It's sort of a catch-22, but it's in the hands of the hardware manufacturers, since MP3 holds the popularity crown and is not likely to lose it without hardware support.
Bring back dynamic range... www.loudnessrace.net

  • Dibrom
  • [*][*][*][*][*]
  • Administrator
Possible Poor Performance Of Alt-presets
Reply #3
Quote
Dibrom,

Are you basically stating here that due to circumstances, it's unlikely the --alt-presets can/will ever be further improved (significantly) in terms of quality?


Unfortunately, yes.  There may be improvements, but unless the core psymodel for LAME itself is significantly improved first, then the --alt-presets aren't going to get a whole lot better than they are now.  And since it seems unlikely that this will happen (at least anytime remotely soon), well.. you get the idea.

Quote
As for the other formats, unless/until they have some significant hardware support they are nearly useless, IMHO (who wants to sit in a hard desk chair in front of a computer, listening to music?).  It's sort of a catch-22, but it's in the hands of the hardware manufacturers, since MP3 holds the popularity crown and is not likely to lose it without hardware support.


You don't have to listen in front of your computer.  Many people on these forums run lossy audio formats through their main sound system in their house.  Of course, you can't take all of this with you in your pocket yet, but you still aren't simply restricted to being physically in front of the desktop.

And there are already some possibilities for portable usage of some other formats.  For example, I can play Ogg Vorbis files on my PDA.  If I spent some time, I could probably even get mppdec working on it.  FAAD2 was apparently integerized recently, so that's another possibility..

Honestly, most of these not-quite-so-mainstream formats will probably never be well supported by the common portable product, but I expect that at some point we will start to see more specialty products pop up which are capable of supporting nearly any format out there.  All we really need is for someone like iriver to release an API and development kit, or something similar.  It's got to happen eventually.  Either that or PDA's will continue to drop in price along with solid state media until it all reaches a point where it's more practical to use something like that then to rely on more hardcoded playback devices.

It's probably likely that the latter will happen sooner than the former, unfortunately, but I guess that's the price to pay for wanting something better -- which is almost always not the mainstream choice.  It always seems to take a little more time, a little more effort, and often (though not always) a little more money to get the best.

  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #4
Would gpsycho be any better in this respect?

Possible Poor Performance Of Alt-presets
Reply #5
Surely not overall, I doubt even a few cases
For those who want better quality that --alt-preset standard, alternatively extreme offers, MP3 is really not your choice...
Try out other formats, MPC, AAC..
Atleast give it a shot

  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #6
Quote
Atleast give it a shot

I haven't tried Vorbis (yet), but I've occasionally tried MPC.

The pity is that, as many people points out, I can't use any of them with my portable iRiver IMP-100/Rio Volt SP100/Soul DMP-01.

I guess that when an Ogg Vorbis decoding engine for this iRiver players is released, bot (Vorbis and iRiver) will receive a big boost.

Possible Poor Performance Of Alt-presets
Reply #7
well. MPC is easily transcoded with MPxChange

  • [JAZ]
  • [*][*][*][*][*]
Possible Poor Performance Of Alt-presets
Reply #8
-api ( insane, CBR320 ) is the best that MP3 can offer (without counting freeformat, that isn't much compatible).

If you hear problems with -api (as well as with fhg at 320kbps), then you can be sure that file just can't sound better (with mp3).
I found at least one file like this, which was discussed some months ago. (search for awe32 in the mp3 forum. the reason of the name is a long story)

In other words: appart from the limitations of the format itself (pre-echo, scalefactor band 21...) mp3 has a bitrate limitation, which in some samples is not enough.

  • Dibrom
  • [*][*][*][*][*]
  • Administrator
Possible Poor Performance Of Alt-presets
Reply #9
Quote
Would gpsycho be any better in this respect?

You mean better in the --alt-presets?

No.  Nspsytune has less problems than GPSYCHO if you compensate for some of the other issues nspsytune has, which the --alt-preset vbr modes do very well.

  • wkw
  • [*][*]
Possible Poor Performance Of Alt-presets
Reply #10
I did some studies on the awe32.wav clip with AAC encoder. It has the same problem as with fatboy.wav. The attacks are too closely spaced and somehow, temporal processing is not sufficient to compensate for pre-echo.. As a result it is necessary to switch to short-block completely just as Ivan has pointed out for fatboy.wav (search for AAC forum about QTime MPEG4 encoder)..



wkw

  • Pio2001
  • [*][*][*][*][*]
  • Global Moderator
Possible Poor Performance Of Alt-presets
Reply #11

  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #12
Well, the thing is that with this trumpet sample there are no pre-echo issues or dropouts, and the the VBR alt-presets don't run out of bits, most frames are 192 Kbps with --aps and ape, only that with --ape there are some more 224 and 320 Kbps frames.

However, --api sounds much better, and I wasn't able to ABX it, whilst the other VBR ones were quiet obvious. So, this suggests that the problem may not be in the LAME psychoacoustic core, but more in the VBR handling?

  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #13
Ehm... any comment on this? 

  • Dibrom
  • [*][*][*][*][*]
  • Administrator
Possible Poor Performance Of Alt-presets
Reply #14
The "VBR handling" and the psychoacoustics are kind of the same thing really.  The fact that the encoder doesn't recognize that these areas may need more bits (that may not be the problem though), would be due to the psymodel not analyzing the samples properly and giving them enough emphasis.

With --api, you are kind of using "brute force", and just kind of waste bits beyond what the psymodel would normally deem as necessary.  So that's probably why that ends up sounding better.

  • JohnV
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #15
Quote
However, --api sounds much better, and I wasn't able to ABX it, whilst the other VBR ones were quiet obvious. So, this suggests that the problem may not be in the LAME psychoacoustic core, but more in the VBR handling?

If VBR doesn't get it right like happens with both gpsycho and nspsytune, then it's directly psychoacoustics' fault. VBR mode determines the number of bits used based on the (meant to be inaudible) quantization noise, which depends directly on the masking threshold created by psychomodel, defining how much and where it allows noise to be introduced.

CBR and ABR are using rougher method when deciding how many bits are needed: perceptual entropy. And with higher CBR bitrate, you have plenty of error margin, which is advantageous compared to VBR when it fails. Of course CBR can run out of bit reservoir and also perceptual entropy can fail in ABR's case (doesnt still allocate enough bits), which also lead to problems, unless you use very high cbr/abr bitrate.

Anyway, this is clearly a deficiency in Lame psychoacoustics.
Also it's a very short clip. Lame VBR often has trouble adjusting for few seconds if the start of the clip is very tricky.. The trumpet sounds a bit better at least with APS if you copy the section to the end, so that there are 2 trumpet sections and encode that. The latter section sounds a bit better imo (only did a very quick test).
Juha Laaksonheimo

  • KikeG
  • [*][*][*][*][*]
  • Developer
Possible Poor Performance Of Alt-presets
Reply #16
Quote
The trumpet sounds a bit better at least with APS if you copy the section to the end, so that there are 2 trumpet sections and encode that. The latter section sounds a bit better imo (only did a very quick test).

In my opinion, when doing so the result is the same, still very easily ABXable.