IPB

Welcome Guest ( Log In | Register )

 
Reply to this topicStart new topic
Low Bitrate VBR sounds worse in 3.98
audiobookie
post Nov 6 2008, 01:06
Post #1





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



I initially thought vbr had improved in 3.98 until I looked at the file size.

This showed up when encoding spoken word audiobooks.

3.98 V8 does sound better than 3.97 V8, but it should it gives about the same file size as 3.97 V7.
Using the new floating point, V9.1 gives the same size as the old V8 and sounds much worse.

I know vbr is more quality orientated, and abr is for a more predictable file size, but it struck me strange.
I thought the old vbr-new, which I was using in 3.97, had now just become the default vbr.
It seems a lot more has changed and not for the better.

The increased size adds up over a 16 hour book, think I'll go back to 3.97.

I would be interested to hear any ideas or opinions on this.

Cheers.
Go to the top of the page
+Quote Post
dyneq
post Nov 6 2008, 04:09
Post #2





Group: Members
Posts: 143
Joined: 29-December 05
Member No.: 26719



Is stereo important to you? If not, give
CODE
-V8 -m m --resample 24
a try. Transparent to me and very small file size.
Go to the top of the page
+Quote Post
audiobookie
post Nov 6 2008, 04:34
Post #3





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



I will give your settings a try, and yes I encode as mono files.

Still very curious as to why the quality/size ratio altered so much.
Go to the top of the page
+Quote Post
euphonic
post Nov 9 2008, 05:57
Post #4





Group: Members
Posts: 86
Joined: 12-February 06
Member No.: 27711



QUOTE (audiobookie @ Nov 5 2008, 20:34) *
I will give your settings a try, and yes I encode as mono files.

Still very curious as to why the quality/size ratio altered so much.


I've noticed that too. I guess part of the reason is because --strictly-enforce-ISO has been defaulted in 3.98.
Go to the top of the page
+Quote Post
Lyx
post Nov 9 2008, 07:56
Post #5





Group: Members
Posts: 3353
Joined: 6-July 03
From: Sachsen (DE)
Member No.: 7609



Actually, it has been known since quite a while that at below 90kbps, ABR tends to give higher quality than VBR. This does of course not invalidate the regression which you are pointing out. Just mentioning that below 90kbps, the "VBR is better"-mantra is no longer true.
Go to the top of the page
+Quote Post
BlAcKnOiSe
post Nov 9 2008, 11:30
Post #6





Group: Members
Posts: 14
Joined: 8-August 08
Member No.: 56869



From the Hydrogenaudio wiki:
QUOTE
For very low bitrates, up to 100kbps, ABR is most often the best solution. Use --abr <bitrate> (e.g. --abr 80).

--preset voice is only available in the command line front-end, and is there for compatibility. It is currently mapped to --abr 56 -mm, so that means that the recommendation would be to encode in mono, and use ABR.

Also note, that far as I know, lame is tuned down to 80 kbit/s only.
Go to the top of the page
+Quote Post
audiobookie
post Nov 10 2008, 02:30
Post #7





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



I did a lot of testing last year, using 3.97 then of course, and read the wiki before settling on V8.

ABR may be better for music at low bitrates, but with audio books they sounded better VBR every time.
It may be because audio books can be very dynamic. A read of todays newpaper stories is very different audio to an actor shouting "RUN FOR COVER" one minute then whispering "i love you babe" next minute as he is about to die for example.

Just did some tests with ABR in 3.98 and still find this to be true, when going for this target quality and size.

@ euphonic, I used ISO enforce in 3.97 too.

This post has been edited by audiobookie: Nov 10 2008, 03:26
Go to the top of the page
+Quote Post
lvqcl
post Nov 10 2008, 11:51
Post #8





Group: Developer
Posts: 3341
Joined: 2-December 07
Member No.: 49183



QUOTE (audiobookie @ Nov 6 2008, 06:34) *
Still very curious as to why the quality/size ratio altered so much.

I encoded quiet piano music with 3.97 -V7 --vbr-new and 3.98.2 -V7. Bitrates are 52 and 72kbps respectively, and spectrograms show that 3.98 keeps more quiet sounds than 3.97 (left is 3.97, right is 3.98). Is it bad or good -- that's quite a different issue.

Go to the top of the page
+Quote Post
audiobookie
post Nov 12 2008, 05:34
Post #9





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



My point entirely. With the increase in kbps in the same V setting shown here, wouldn't you expect a better quality?

@ lvqcl, I'd be interested to know how 3.97 at -V6 looked and sounded to you.
It could be just like 3.98 -V7!
Go to the top of the page
+Quote Post
jetpower
post Nov 12 2008, 14:47
Post #10





Group: Members
Posts: 67
Joined: 6-September 05
Member No.: 24363



As audiobook are listened at moderate volume this keeping of very quiet sounds IMHO mostly wastes bitrate here. These quiet sounds are usually noise and/or not important really as they are not typically heard anyways under casual listening.

If you raise ATH in LAME by 15 dB (--athlower -15) you may notice no to little difference at a gain of 15-25% lower bitrate IIRC (at stereo, 44.1 khz). Say from 100kbps to 80 kbps. I used even -25 dB and this seemed good trade-off for my audiobook listening purposes.
However --athlower switch does not work in normal LAME builds. I wish it did because the debug build I've got is a bit slow.
Go to the top of the page
+Quote Post
audiobookie
post Nov 13 2008, 01:33
Post #11





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



Jetpower, are you saying --athlower - 15 won't work in 3.97? Shame I'd like to try that.

Here is an example of what I found with the two versions, using a short edit of 44.1 kHz mono.

3.98: V9.1 Size 405.9 kb and 51 kbps.
3.97: V8 Size 400.1 kb and 50 kbps.

The 3.98 is noticably worse, although size and kbps are virtually equal.

I know I'm using setting not commonly used, the current poll on what people use shows that.
But if this result extends to higher quality V settings as well, what has really improved with 3.98?
Go to the top of the page
+Quote Post
jetpower
post Nov 13 2008, 16:24
Post #12





Group: Members
Posts: 67
Joined: 6-September 05
Member No.: 24363



I had this LAME linked in this post .
It has those weird experimental options enabled:) - but is a third slower than rarewares LAME.
Now that I tried it again it seems you can even go as low as -40dB and at the same time adjust -V quality to get good bitrate.

This post has been edited by jetpower: Nov 13 2008, 16:24
Go to the top of the page
+Quote Post
[JAZ]
post Nov 13 2008, 20:25
Post #13





Group: Members
Posts: 1767
Joined: 24-June 02
From: Catalunya(Spain)
Member No.: 2383



QUOTE (audiobookie @ Nov 13 2008, 01:33) *
3.98: V9.1 Size 405.9 kb and 51 kbps.
3.97: V8 Size 400.1 kb and 50 kbps.


Aside of the fact that two files with equal size will have equal kbps (since one is a reflection of the other, not counting metadata like an ID3 tag), you are telling the encoder to do different things in those commandlines.

Let me explain it:

It is a known fact that lame 3.98 has increased the average bitrate of the -V settings. Am I saying that it needs more bitrate to do the same? Yes, almost. The tweakings it has received are affecting the whole range of the -V setting. For the same setting, it looks as an improvement in quality, but for the same bitrate (as you found out) especially on the highest and lowest ends, the results translate to more than one point of the -V setting, and as such, applying lower quality setting parameters.

Of course, with a content like voice, it makes sense to use VBR, since the encoder could adapt to the demands, keeping the bitrate quite low. ABR does not help too much, since the ABR algorithm, at low bitrates, will not switch much far from the target bitrate, and as such not really adapting to quality demands.

The only thing that i can tell you is that currently 3.97 fits your scenario better, and that no effort of tweaking 3.98 will give a similar result.

Is this going to be addressed? I can't tell for sure. It sure there is enough difference for them to check what has happened (requiring 25% more of bitrate for a similar quality is just too much), and probably part of the problem is shown in that screenshot that lvqcl posted. (Screenshots do not show quality, but can show defects), which is precisely the effect that the setting jetpower mentioned has.


QUOTE (jetpower @ Nov 13 2008, 16:24) *
It has those weird experimental options enabled:) - but is a third slower than rarewares LAME.
Now that I tried it again it seems you can even go as low as -40dB and at the same time adjust -V quality to get good bitrate.


Those settings are just in the debug builds, since they can be used, at the indication of the developers, to know if an internal parameter should be better adjusted to another value, or if it has relation to an artifact.

Free experimentation of those settings is not encouraged, nor is it recommended. It is especially important to notice that where a setting can show no difference to an encoded piece, it can change another quite much.

This post has been edited by [JAZ]: Nov 13 2008, 20:27
Go to the top of the page
+Quote Post
jetpower
post Nov 13 2008, 21:29
Post #14





Group: Members
Posts: 67
Joined: 6-September 05
Member No.: 24363



LAME here is overly careful with ATH because it cannot foresee listening conditions. As the mp3 will be for personal use I see no problem with using experimental switches and this one in particular. I think the goal here is not transparency but optimal compression - to cram as much voice recording as possible with acceptable quality on the device one uses.

This post has been edited by jetpower: Nov 13 2008, 21:30
Go to the top of the page
+Quote Post
audiobookie
post Nov 13 2008, 23:48
Post #15





Group: Members
Posts: 22
Joined: 16-January 07
Member No.: 39699



Jaz, thanks for the detailed answer.

When 3.98 was officially released it was stated vbr-new was now the default. Obviously the experimentation with vbr-new has gone on beyond what it was in 3.97, rather than just vbr-new as it was then being the new default. I can see how the changes have come about.

I will stick with 3.97 for the moment, it is more efficient for what I'm doing.
Go to the top of the page
+Quote Post
Dynamic
post Nov 14 2008, 01:45
Post #16





Group: Members
Posts: 796
Joined: 17-September 06
Member No.: 35307



QUOTE (audiobookie @ Nov 13 2008, 01:33) *
3.98: V9.1 Size 405.9 kb and 51 kbps.
3.97: V8 Size 400.1 kb and 50 kbps.

The 3.98 is noticably worse, although size and kbps are virtually equal.


According to the LAME Wiki V9.1 would be at 24 kHz sampling rate while V8 is at 32 kHz sampling rate. The lowpass frequency is also greatly reduced with V9.1. Such a step change is a major difference when you maintain constant bitrate in this way.

Your experience indicates that the bang-for-your-buck (or bang-for-your-bitrate) tradeoff in LAME 3.98.2 has worsened for speech in the low quality region. It's plausible that music will have fewer annoying artifacts and might be improved despite the loss of higher frequencies (which is at least a consistent degradation) when you target the same bitrate, or it might be that perceived quality per bitrate in non-transparent regions has suffered from changes that maintain transparency better in the higher-quality -V settings.
Go to the top of the page
+Quote Post
lvqcl
post Nov 14 2008, 02:21
Post #17





Group: Developer
Posts: 3341
Joined: 2-December 07
Member No.: 49183



QUOTE
According to the LAME Wiki V9.1 would be at 24 kHz sampling rate


3.97 (V9) - yes, but 3.98.2 - 22.05 kHz. (tested just now).

This post has been edited by lvqcl: Nov 14 2008, 02:23
Go to the top of the page
+Quote Post
robert
post Nov 16 2008, 14:18
Post #18


LAME developer


Group: Developer
Posts: 788
Joined: 22-September 01
Member No.: 5



QUOTE (lvqcl @ Nov 10 2008, 11:51) *
QUOTE (audiobookie @ Nov 6 2008, 06:34) *

Still very curious as to why the quality/size ratio altered so much.

I encoded quiet piano music with 3.97 -V7 --vbr-new and 3.98.2 -V7. Bitrates are 52 and 72kbps respectively, and spectrograms show that 3.98 keeps more quiet sounds than 3.97 (left is 3.97, right is 3.98). Is it bad or good -- that's quite a different issue.

What you are seeing is the workaround for the "sandpaper noise problem" at work. We may revisit it with 3.99 and maybe poeple interested at low quality settings take a look at beta versions then.
Go to the top of the page
+Quote Post
DigiWiz
post Nov 23 2008, 02:24
Post #19





Group: Members
Posts: 1
Joined: 22-November 08
Member No.: 62936



I had been using Razorlame and LAME 3.97b2 combo for many years, and was very happy - but when I saw 3.98.2 I thought, "Whoa, I'm way behind the times." So I installed and did some tests.

Most of my LAME transcoding involves speech - specifically AM radio broadcasts, which are clearly recorded by people who don't know what they're doing with LAME. One show in particular, is typically 55MB in size.

I then retranscode using my aforementioned combo set using -V 9 --vbr-new, which gives me great, nearly transparent, if not transparent quality, and the resultant filesize is ~22MB-24MB, a substantial savings.

Last night I used Razorlame with the latest 3.98.2 and was surprised beyond belief... using -v9 as the setting, the filesize ballooned upwards by 43%!

So I took a "War of the Worlds" radio broadcast I have as an mp3, and expanded it to a wave file: 306791KB (per Windows explorer). Using 3.97b2, resultant mp3 filesize was 8822KB, and sounded like all my other AM radio transcodings - virtually transparent. So I tried Razorlame with 3.98.2. First, as most may have read, the bit-rate bars do not function at all, and the progress bars do not function as well - stay at 0%. After transcoding was finished, I was astounded to see the -v9 3.98.2 filesize had ballooned to 12654KB, again, a 43% increase in size!

I have not tried 3.98.2 with music yet, but it will be interesting - I usually use -v2 for most stuff, so will do a few tests to see if the filesize differences are noticeable. I can't imagine real-world listening experience would change perceptively enough for a human to notice.

I'll have to say, a 43% increase in filesize for same-setting, low-grade speech/AM radio broadcast material sucks donkey-balls.

DW

Windows XP sp3
Go to the top of the page
+Quote Post
pannayar
post Feb 7 2009, 08:56
Post #20





Group: Members
Posts: 1
Joined: 7-February 09
Member No.: 66706



I have been using 3.97 and recently thought oh! there is the new 3.98 for me.... I have been encoding a lot of audio books lately and have been disappointed with the inreased file sizes from 3.98. I am sticking to 3.97

For audio books I do
-mm --vbr-new -V9 --lowpass 9.5

When V9 makes the quality drop noticeable and I switch to V8...
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: 1st August 2014 - 23:25