Update: better pre-echo for medium/standard/extreme

New presets: portable/portable1 (targetting 130kbps on average)

Win32 binaries available at 'Other Stuff' Mirror 1.

I did a few quick ABX tests using the 41_30.wav test sample encoded at APS on the new v3.94 A-12 codec. I was able to ABX v3.90.2 from v3.94 Alpha 12 15/15 times, with 3.90.2 having the better quality imo. Also able to ABX v3.94 a12 from the original 25/25 times, and this was EXTREAMLY easy to do (Took about 2 min). The first 1.5 sec of the A12 clip are washed out, with a sort of warble. The better quality of v3.90.2 could be due to the much higher bitrate as opposed to the new Alpha 12 (224 vs. 190 kbps).

**************************************************************************************

[LAME v3.90.2 --alt-preset standard || 41_30.wav Sample]

Bitrates:
----------------------------------------------------
128 | 1.1%
160 |||||||||||||| 13.7%
192 |||||||||||||||||||||||||||||||||||||||| 36.7%
224 |||||||||||||||||||||| 20.8%
256 ||||||| 7.0%
320 |||||||||||||||||||||| 20.6%
----------------------------------------------------

Type : mpeg 1 layer III
Bitrate : 224
Mode : joint stereo
Frequency : 44100 Hz
Frames : 1151
ID3v2 Size : 0
First Frame Pos : 0
Length : 00:00:30
Max. Reservoir : 208
Av. Reservoir : 71
Emphasis : none
Scalefac : 0.3%
Bad Last Frame : no
Encoder : Lame 3.90

Lame Header:

Quality : 78
Version String : Lame 3.90
Tag Revision : 0
VBR Method : vbr-old / vbr-rh
Lowpass Filter : 19000
Psycho-acoustic Model : nspsytune
Safe Joint Stereo : yes
nogap (continued) : no
nogap (continuation) : no
ATH Type : 4
ABR Bitrate : Unknown
Noise Shaping : 2
Stereo Mode : Joint Stereo
Unwise Settings Used : no
Input Frequency : 44.1kHz

**************************************************************************************

[LAME v3.94 Alpha 12 --preset standard || 41_30.wav Sample]

Bitrates:
----------------------------------------------------
96 0.3%
112 0.6%
128 |||| 4.5%
160 |||||||||||||||||||||||||||||||||||||||| 41.6%
192 ||||||||||||||||||||||||||| 28.6%
224 |||||| 6.3%
256 |||||||||||||| 15.0%
320 || 3.0%
----------------------------------------------------

Type : mpeg 1 layer III
Bitrate : 190
Mode : joint stereo
Frequency : 44100 Hz
Frames : 1151
ID3v2 Size : 0
First Frame Pos : 0
Length : 00:00:30
Max. Reservoir : 215
Av. Reservoir : 59
Emphasis : none
Scalefac : 1.5%
Bad Last Frame : no
Encoder : Lame 3.94 (alpha)

Lame Header:

Quality : 57
Version String : Lame 3.94 (alpha)
Tag Revision : 0
VBR Method : vbr-old / vbr-rh
Lowpass Filter : 19000
Psycho-acoustic Model : nspsytune
Safe Joint Stereo : yes
nogap (continued) : no
nogap (continuation) : no
ATH Type : 4
ABR Bitrate : Unknown
Noise Shaping : 1
Stereo Mode : Joint Stereo
Unwise Settings Used : no
Input Frequency : 44.1kHz

Edit - It's 41_30.wav, not 30_41.wav as JohnV pointed out .

I did a very quick abchr of castanets.wav:

bitrates
3.90.2 --alt-preset standard: 207
3.94a12 --alt-preset standard: 211
3.94a12 --alt-preset medium: 170
3.94a12 --alt-preset 170 --scale 1: 181

abchr ratings:
medium: 3.6
abr 170: 4.1
3.94 aps: 4.5
3.90.2 aps: 4.7

Note: The 3.6 rating isn't really as bad as it seems: I spread out the scale when rating. Rating is based mostly on the first slow castanet pre-echo.

ff123

CASTANETS.WAV

3.90.2 = 207 kbps
3.94.a12 = 211 kbps
Both are easy to differenciate from original, but I can't decide which is better.

A single trumpet note from Miles Davis

3.90.2 = 189 kbps --- note = 3.5 / 5
3.94.a12 = 166 kbps --- note = 1 / 5
Awful distorsion with 3.94. I can't support it...

some harpsichord notes

3.90.2 = 194 kbps --- note = 2.8 / 5
3.94.a12 = 185 kbps --- note = 3.7 / 5
Both are distorded, as usual with mp3... but 3.90.2 had more pre-echo -> easiest to ABX

erhu sample

3.90.2 = 168 kbps --- note = 2 / 5
3.94.a12 = 180 kbps --- note = 4.7 / 5
Impressing bugfix. 12/20 on a quick ABX test with alpha12. Very hard to hear a difference in a noisy environment (CPU fan)

Jump sample

3.90.2 = 238 kbps --- note = 4.8 / 5
3.94.a12 = 207 kbps --- note = 4.3 / 5
HF Noise with alpha12. 3.90.2 is near perfection for me. Both are really good.



The best pre-echo progress I can surely hear is on the harpsichord sample. alpha12 is correcting the (rare) bug on the erhu sample. Nevertheless, the Miles Davis trumpet distorsion really worries me.

bounce_clip2 results:

3.90.2 --alt-preset standard: 206 kbit/s, 4.6
3.94a12 --alt-preset standard: 192 kbit/s, 4.3
3.94a12 --alt-preset medium: 157 kbit/s, 3.9
3.94a12 --alt-preset 170 --scale 1: 188 kbit/s, 3.6

I chose a section near the beginning and listened for pre-echo.
3.90.2 aps is definitely the best for the small section I listened to (abx'd 8/8 vs. 3.94a12 aps).

ff123

Edit: again: pay no attention to the wide ratings differences. The audible differences are actually quite small.

ff123: not very important, but I am targetting 165kbps for medium.

guruboolez: is the catastrophic trumpet worst than with 3.94a11?

I searched in the forum, and the real catastrophic sound occurs with alpha9 --preset standard4 (this topic).
I've uploaded a short sample, http://membres.lycos.fr/guruboolez/AUDIO/s...Dance_short.ape


I tried to compare the quality of the latest alpha on this sample :

3.90.2 = 4.7/5 (extremly good ; I had to listen to it many times in order to find a small difference)
3.94.a9 --standard4 = 1/5
3.94.a10 = 2.5/5
3.94.a11 = 2.5/5
3.94.a12 = 2.0/5

I can't be sure that a.12 is worse than a.11 & a.10 : they are really close.
Nevertheless, the historic preset is clearly transparent, when the 3.94 series isn't (distorsion)

I know this question must sound naive, but why is is LAME 3.90.2
so hard to improve on? From what I've read, newer LAME versions
not only fail to improve on 3.90.2, but perform much worse in listening
tests. I would have thought that the only way is up...

Well, I'm not sure if the current 3.94 are really worst than 3.90.2.
The discussions are centered on bad results, because from a developement point of view it is what is interesting and what we should focus on.

So here is THE question: overall, is the current 3.94 better/equal/worst than 3.90.2?
It would be interesting to have opinions about this.

Gabriel > I checked again the space issue with 3.94 standard series and noisy tracks.




** 99,5 % of 128 kbps frames with 3.90.2 || wide distribution with 3.94 (112 = 14% , 128 = 48%, 160 = 34%)
During of tracks : between 120 and 300 seconds

· Total (6 files) :
3.90.2 = 20.2 MB | 173.3 kbps
3.94.a = 21.5 MB | 185 kbps
=> + 6.75 % (+12 kbps)

I compared 3.90.2 and 3.94.a12 --standard profile on a contempory piano track. Nice and clear recording. I'm impressed by the overall quality of the new compile. On the sharpest attacks, 3.90.2 is easy to ABX (pre-echo, smeared attacks). 3.94a12 is much harder : crisp sound.

EDIT both are 137 kbps (30 seconds). I can't put encspot graphs, but they are interesting. 3.94 is generous in 256 & 320 frames in the pre-echo area. Bitrate is the same, because of 37% of 96 & 122 kbps frames. 3.90.2 is giving 80% of 128 kbps frames.

I've done some tests with a new sample: campestre.flac

Testname: campestre Listening Test

1R = abr165_3902.wav
2L = apma12.wav
3R = abr165_a12.wav

---------------------------------------
General Comments:

apma12.wav = 3.94 alpha 12 (--preset medium), 211 kbps
abr165_3902.wav = Dibrom's 3.90.2 (--alt-preset 165 --scale 1), 166 Kbps
abr165_a12.wav = 3.94 alpha 12 (--preset 165 --scale 1), 169 Kbps

All ratings based listening the whole sample, no need to ABX
---------------------------------------
1R File: abr165_3902.wav
1R Rating: 2.5
1R Comment: noisy, especially in range 11.8 - 14.1
---------------------------------------
2L File: apma12.wav
2L Rating: 3
2L Comment: Better on difficult parts but noise is still present at 11.8 - 14.1
---------------------------------------
3R File: abr165_a12.wav
3R Rating: 1.0
3R Comment: Very annoying, a lot of spread noise and chirping. Possible 3.94 alpha ABR flaw ?
---------------------------------------

Note: Encspot says that abr165_a12.mp3 has a lot of ss frames if compared with abr165_3902.mp3

***************************************


Testname: campestre Listening Test

1R = apsa12.wav
2L = 3902aps.wav

---------------------------------------
General Comments:

aps3902.wav = Dibrom's 3.90.2 (--alt-preset standard), 246 Kbps
apsa12 = 3.94 alpha 12 (--preset standard), 254 kbps

All ratings based listening 11.8 - 14.1 range
---------------------------------------

1R File: apsa12.wav
1R Rating: 3.0
1R Comment: noisy and chirping
---------------------------------------
2L File: 3902aps.wav
2L Rating: 4.3
2L Comment: noisy but clearly better than 1R
---------------------------------------

ABX Results:

Original vs aps3902.wav
10 out of 10, pval < 0.001
Original vs apsa12.wav
10 out of 10, pval < 0.001
apsa12.wav vs aps3902.wav
10 out of 10, pval < 0.001

I'm confused about the output for this test. On top it says that 1L is aps3902, but on the bottom it says that 1R is aps3902.

What happened?

ff123

I've done a new quick test and corrected the previous post. Thanks ff123.
maybe I made a cut&paste error when renaming the filenames in a more comprehensive manner.

I wouldn't use --scale 1. I know it will cause problems with ABR even at 165kbps. If you want to keep everything at the same volume, try using --scale 0.95 with all three files.

This afternoon i made some more tests with different samples and the problem seem not to be an isolated case. I'm not skilful to describe artifacts but the sound is like chirping/waving and the problem is still present even with CBR.

I've uploaded some encodings (--alt-preset 128 --scale 1) to allow a quick comparision between 3.90.2 and 3.94a12:

applaud.zip
liebestod.zip
ma11.zip
avril_lg_clip_.zip
love.zip (includes alpha11 encodes)

Artifacts are *very* easy to hear with applaud and love sample.
I've noticed that the problem is still present in alpha 11 (19 Feb) and not in alpha11 (8 Feb) both compiled by John33.

Anybody can confirm this ?

As edited on the first post, there are now portable/portable1 presets.

I'd also like more reports about this potential abr flaw. Could anyone confirm/infirm that something appeared between 8 and 19 feb?

I hate to sound like a broken record, but do not use --scale 1 with ABR or CBR. There is a "bug" with LAME that requires the use of scaling with these encoding modes at low and moderate bitrates (<240kbps). For instance, --preset cbr 128 automatically uses --scale 0.93...do not override this with a higher value. If you want to keep the volume the same across files for ABX comparisons, use --scale 0.93 with all your (testing) commandlines.

It's entirely possible that the chirping/waving problem you describe is caused by using --scale 1. Try --scale 0.93 instead and see if the problem persists.

More info on this 'bug'? I've encoded alot of alt-preset 128 --scale 1 with 3.90.2 and not noticed anything - though I use mp3gain on them after encoding.

Apr 1. joke by mithrandir? The --scale 0.93 is there to decrease possible clipping artifacting. I really don't think that simply scaling the output to --scale 1 can cause any "chirping/waving problem" (I don't even remember hearing that claim before), but it can result more clipping which you can indeed handle with mp3gain for example.
Maybe the clipping can sound sometimes like "chirping", but if the problem is clipping, we should talk about clipping, not about "chirping/waving" which would indicate psychoacoustic problems. Never heard that clipping could sound like "waving" though.

As JohnV already said, scaling should not be responsable for chirping/waving problems. At least in one case (campestre.wav) the signal does not clip even using --scale 1. Scaling in *every cases* is not necessary. Moreover, i'm comparing two encoded files (3.90.2 vs 3.94a12) both obtained using --scale1 and ONLY the 3.94a12 encoded has such problems.

I'm still waiting for a confirm of these problems that are quite easy to detect, try with love.wav or applaud.wav sample. Thanks.

@floyd, the problem does not affect v3.90.2

Clipping, chirping, waving or whatever you want to call it - basically it's something that shouldn't be there. I'm reasoning that if somebody hears something wrong with alpha12 using --scale 1 it could be because of the lack of scaling, not necessarily because there is a unique alpha12 problem.

good question...

I'm interested in the portable and medium presets for my flash based portable. I was just playing around with them and these are the results I got:

3.90.2 APS -Y: 181 kbps
3.94a12 medium: 193 kbps
3.94a12 portable: 162 kbps
3.94a12 portable1: 179 kbps
3.94a12 preset 160: 157 kbps

I used Prayer by Disturbed. medium ended up bigger than APS -Y? portable had very bad artifacting, which I ABXed 5/5. portable1 and even preset 160 were noticably better than portable. I can provide clips if someone can give me somewhere to upload them to.

Edit: actually 3.94a12 APS -Y ended up at 204 kbps so medium was a little smaller.

Obviously something weird happened.
I checked --alt-preset 128 Applaud.wav and there's definitely something wrong with a12 and some earlier alpha builds. There's loud watery high frequency artifacting most audible during 3s and 6-8sec.
Also, check these spectras:

Lame 3.90.2 --alt-preset 128 (120.9 kbps)
Mar 31. Lame 3.94a12 --preset 128 (131.1 kbps)

a10 is better, with somewhat similar spectra as 3.90.2, although a10 is not quite as good as 3.90.2 because of some dropouts like can be also seen from its spectra (6-8sec):
Feb 3. Lame 3.94a10 --preset 128 (127 kbps)

Something wrong probably slipped into a12 for cbr/abr.
We know the (probable) cause, and Takehiro will have a look at it.

Back to preset standard.
I encoded a track, live recording, and I expected a consequent bitrate difference between 3.90.2 & 3.94a12. I was surprised to see that the alpha version is really smaller than the 3.90.2 one.
On the whole track (5 minutes) :


On the short sample I uploaded (flac) :


The size is nice, but the quality is worse than the historic preset.

Greetings LAME devs,

Is there likely to be any improvements to the quality of 320 kbps CBR in the LAME releases after 3.90.2 or is work going mainly into the determination of which bitrates/settings to use for each frame to optimise (i.e. reduce the overall bitrate of) the preset VBR modes while not losing quality in them?

Basically, have we hit the wall when it comes to the SQ possible with 320 CBR? Is there anything left to try? Is anything being attempted with respect to this?

RD.

I tryed the Ranou sample http://membres.lycos.fr/guruboolez/AUDIO/s...les/Rannou.flac
This sample is 160kbps with the current preset standard and was 204 with 3.90.2

I tryed to abx it, but only reached 4 out of 10. How much do you score on this sample with 3.94 and 3.90.2?

One particular thing about this sample is that it's nearly not dependant on the ath even despite its very low volume.
I tryed --noath and the bitrate was similar. I also tryed --scale 3 (the value is high but there was no clipping) and the bitrate was also similar.

The score I did two days ago (quick test, on seconds 2.5-5.0) :

I hope the new lame release doesn take too much longer...80% of the new songs are encoded with Lame 3.93...

How do you know that they are encoded with 3.93 and not 3.93.1?
I think that most of the "download locations" only have 3.93.1 and not 3.93

Well, I do not know it for sure, since EncSpot doesn't make a difference on that...but it doesn't matter because didn't you say that all the people should stick with 3.92 if they didn't have any problems with this release? Usually what people do is downloading the newest release, because they suppose that this is better than the previous release (me too). I'm only afraid that most of those release groups won't upgrade that fast (when a new lame release is finally out) since they have already upgraded to 3.93 .
But I know that you are doing your best . Is there even something more to do? I mean when it takes so much effort to improve lame I would assume that this (lame) technology has reached its limits.