Skip to main content
Topic: LAME 3.99 is out (Read 219443 times) previous topic - next topic
0 Members and 2 Guests are viewing this topic.

LAME 3.99 is out

Reply #1
Congratulations to developers of LAME with their new release 3.99.

Thank You for keeping the development of high quality  and open source MP3 encoder.

So what will be next? 3.100?

LAME 3.99 is out

Reply #2
Thanks a lot for the new version.

Tuning on VBR scale / resulting bitrate  is much appreciated, as there was a large gap between -V0 and -b 320 - those settings which are generally expected to yield the highest quality.

I'm curious about the tuning on PSY model.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #3
x86 (
LAME 3.99 32bits (
CPU features: MMX (ASM used), SSE (ASM used), SSE2

x64 (
LAME 3.99 64bits (
CPU features: , SSE (ASM used), SSE2

LAME 3.99 is out

Reply #4
So is the best plan to wait for a few more .1 revisions? I heard some complaints about artifacts and buggy psy model in 3.99 beta, have those been fixed? There's nothing in the changelog that would make one think so...

LAME 3.99 is out

Reply #5
For learning about 3.99's average bitrate for -Vx i encoded my usual test set of various pop music:

-V5:  3.98.4: 136 kbps  3.99: 126 kbps
-V4:  3.98.4: 152 kbps  3.99: 147 kbps
-V3:  3.98.4: 166 kbps  3.99: 167 kbps
-V2:  3.98.4: 188 kbps  3.99: 191 kbps
-V1:  3.98.4: 207 kbps  3.99: 224 kbps
-V0:  3.98.4: 233 kbps  3.99: 258 kbps

I like this behavior of 3.99.
For the many -V3 and -V2 users things don't really change.
From -V2 to -V0 it's adequate to increase the average bitrate steps from one quality level to the next because audible differences can only be expected with a higher amount of bitrate difference than in the range below -V2.
-V5 now matches pretty well 128 kbps, which is roughly the expected average bitrate for -V5. This is especially welcome for 128 kbps listening tests.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #6
My test set:

Code: [Select]
         3.98.4       3.99
-V6:     126 kbps     116 kbps
-V5:     139 kbps     131 kbps
-V4.999: 140 kbps     138 kbps
-V4:     155 kbps     151 kbps
-V3:     169 kbps     171 kbps
-V2.999: 179 kbps     182 kbps
-V2:     200 kbps     197 kbps
-V1:     222 kbps     232 kbps
-V0:     252 kbps     270 kbps

-V2 -Y:  181 kbps     185 kbps
-V1 -Y:  197 kbps     214 kbps
-V0 -Y:  220 kbps     242 kbps

LAME 3.99 is out

Reply #7
A big Thank You to the LAME devs for the important work they're doing! (as you can see by my username, I'm a big fan)

An open source, high-quality MP3 encoder continually developed is in my opinion very important, because it still is the de-facto standard for most people when it comes to music files, and it keeps users independent from commercial companies.

I want to donate some money to the LAME project, but I couldn't find a "donate"-button on the site. Does anyone know where it is (if there is one)?


LAME 3.99 is out

Reply #8
I did a short listening test comparing 3.99 against 3.98.4.

I chose samples which are real ugly with CBR 128 (better, but still very bad with -V5):

harp40_1 (harpsichord is a problem for mp3 - I concentrate on the chords around sec. 10.0)
lead-voice (a strong tremolo at sec. 0.0-2.0)
eig (strong pre-echo, with a special problem at around sec. 3.0)
herding_calls (at sec. 0.9-3.7 an artifact is audible)
trumpet (an artifact / an inaccuracy which sounds a bit like that in herding_calls)

The samples with all the encoded mp3 variants can be downloaded from here.

I started comparing -V5.
The result can be summed up quickly: it's all bad, for none of the samples did I have a preference for one of the versions. From that the good message is: 3.99 behaves like 3.98.4, in spite of a somewhat lower average bitrate.

I continued with -V2 and found some differences.
The problem spot with eig at sec. 3.0 is more pronounced with 3.99. On the other hand herding_calls was harder to ABX with 3.99 than with 3.98.4. I was astonished about lead-voice and trumpet: lead-voice because it was so extremely easy to ABX (with both versions), and trumpet because it was rather hard (with both versions).

I finished with -V0.
eig is nearly fine for me as I'm not sensitive for pre-echo. It's just the spot at sec. 3.0 which has me ABX eig rather easily. Again the problem is more pronounced with 3.99. I am also very content with harp40_1 which wasn't easy to ABX (harder with 3.99). The difference with herding_calls was rather strong: With volume set up the right way for this sample (not too loud!) ABXing with 3.98.4 was rather easy, with 3.99 it was very noticeably harder. I ABXed lead-voice easily with both versions. trumpet was hard for me, and with the restricted time I allowed for the test, I could not get sufficiently good ABX results for 3.98.4. With 3.99 I succeeded, but it was very hard, and I consider the 3.99 result to be fine, too. I will redo trumpet tomorrow morning when I hopefully feel fresher than I just did.

As a result I see a certain progress with 3.99. The spot at sec. 3.0 of eig shouldn't be overestimated. For judging pre-echo behavior we hopefully get other listening test results from members who are more concerned about pre-echo problems than I am.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #9
Interesting report.

As a result I see a certain progress with 3.99. The spot at sec. 3.0 of eig shouldn't be overestimated.

I saw You have tested a short time (4 sec) of eig sample.  That's why You have found only one pulsive artifact. The artifacts are more clear on longer version of the sample because it's now the whole train of them.

I found that 3.99 does better than 3.98.4 for eig in my last report.

P.S. But if it was me you will probably saying "There must have gone something wrong"  . You see we are all different listeners, there is no "wrong".

LAME 3.99 is out

Reply #10
Sorry for having said "There must have gone something wrong" about your first report you gave on trumpet behavior. My remark was for the report as such, especially as I missed one result, and I was afraid that you mixed up your 3.97 result with that of a later version. It was not meant to be about your listening experience. When you explained this it's true I was astonished that you preferred 3.97 over 3.98. You're right, we're all different listeners, and our listening experience can vary under different circumstances.

As for eig probably everybody is more adequate to report on pre-echo behavior than I am. That's why I hoped I made it clear that I can only report differences about the spot at sec. 3.0. As for this spot though I'd be surprised if somebody would prefer 3.99. It is possible of course because as you said: we're all different listeners.

I'll have a look at the entire eig sample. I shortened it a long time ago because this allowed me to crank up volume. Later in the track the 'music' gets very loud which prevented me from listening very loud in the first 4 seconds, and with my extract I thought I had the important impulse part. I'll have a look into it.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #11
It's ok  , halb27. My sarcasm sometimes drives me crazy.

Now it makes me wonder what other people hear on eig sample.

LAME 3.99 is out

Reply #12
I didn´t really test anything because i don´t use mp3 much anymore. I encoded my old samples suite with V2 and found no regression on these.
I hear eig with 398.4 and V2 not worse in the whole sample, at least the 15second sample i have. The most obvious difference is the clear artifact at ~second 3 and 9 with 3.99 but me also isn´t to good on typical pre-echo samples.

Edit: Around second 9 both add some artifact that is louder to me with 3.99 also. Besides that i can´t listen such music for to long, so it doesn´t really matter
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME 3.99 is out

Reply #13
I just retested trumpet, and this morning it was not extremely hard to ABX the 3.98.4 result. Today ABXing 3.99 was harder for me. The very effect you ran upon, IgorC.
As both results are so good to me that in practical listening situations with no reference to the original I really wouldn't complain, I'd call both results fine.

As for eig I tried the full sample, and became quickly aware again why I shortened the track for my purpose. The bass after sec. 4 is so strong that I can't hear any issue, not even at around sec. 9.
The discussion on eig made me one thing very clear: I will not report on pre-echo behavior again. I'm well aware now that I used eig in a kind of 'oh, me too can hear some pre-echo problem', but if I can only report about a specific spot within a track full of pre-echo prone impulses, I better am quiet. Guess I will still listen to that spot and possibly report about it, but in the sense of a specific artifact, not in the sense of pre-echo.
As for 3.99 V0's pre-echo behavior I expect it to be better than that of 3.98.4 because 3.99 uses a frame packaging strategy with the target of avoiding a large percentage of otherwise inaccurately encoded frames, and because 3.99 -V0 has higher accuracy demands than has 3.98.4 -V0.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #14
I don't know if this is the right place to ask this question, but I am wondering why LAME 3.99 32-bits and LAME 3.99 64-bits make mp3's that are different in size.
One example is this: an mp3 made with LAME 3.99 32-bits is 7,76 MB (8 138 854 bytes) in size, but the same song encoded with LAME 3.99 32-bits is 7,75 MB (8 137 291 bytes) in size.
The song was ripped with EAC 10. B3 and not modified in any way.
But, I am puzzled about that 1563 bytes in filesize difference

Edit: I forgot to mention that I used LAME 3.99 Bundle compiled with Intel Compiler 11.1. and LAME 3.99 64bit Bundle compiled with Intel Compiler 11.1. Downloaded from rarewares yesterday.

LAME 3.99 is out

Reply #15
Quite simply, it's the fact that one is generating a compile for 32 bit using the internal 'nasm routines' whereas the 64 bit compile uses the compiler's internal processor optimisations. In this case the same compiler is used, but if compiles for the same target are created using different compilers, the encoded output will differ in size to a small extent because the internal math routines differ, amongst other things. I hasten to add that no one, to the best of my knowledge, has yet been able to discern any audible differences between any of the differing outputs generated by the encoders created with different compilers.
My compiles and utilities are at

LAME 3.99 is out

Reply #16
Thank you for replying john33 and thanks for clarifying and explaining the filesize difference for me (and others)

LAME 3.99 is out

Reply #17
Hello. I have a problem with new version.
After upgrading to version 3.99 id3v2.3|id3v1 tags are automatically created during WAV encoding process, showing field <ENCODING SETTINGS> : LAME 32bits version 3.99 ( in foobar2000 v1.1.8.
My LAME parameters in foobar2000 are: --silent --noreplaygain -V0 - %d. With EXACTLY the same parameters id3 tags are not created with LAME 3.98.4.
How to prevent LAME 3.99 from creating id3 tags? I was experimenting with parameters, reading LAME documentation, but couldn't find solution.
iRiver H10 20 GB - Rockbox

LAME 3.99 is out

Reply #18
Confirmed, fixed sources are available from

LAME 3.99 is out

Reply #19
What is everyone's thoughts on -V 0 now using the full spectrum and no low pass filter? Is there really any reason to preserve frequencies above 20 kHz? Would manually applying a low pass filter be wise? I'm also wondering why CBR 320 encodes are not using full spectrum too considering it now uses the same psymodel as VBR.

LAME 3.99 is out

Reply #20
Thanks for the info.
I don't like having no or an extremely high low-pass. It's probably a concession towards those people who come up here occasionally demanding for the settings which allow for the 'full range'.

But it's not a problem as we can use --lowpass. Maybe it's not bad to find out for ourselves where to put the lowpass.
lame3995o -Q1.7
opus --bitrate 140

LAME 3.99 is out

Reply #22
What is everyone's thoughts on -V 0 now using the full spectrum and no low pass filter? Is there really any reason to preserve frequencies above 20 kHz? Would manually applying a low pass filter be wise? I'm also wondering why CBR 320 encodes are not using full spectrum too considering it now uses the same psymodel as VBR.

Very good question. Any explanation for this kind of inconsistency? It seems completely illogical to me to have a lowpass at CBR 320 and no lowpass at -V0

LAME 3.99 is out

Reply #23
Tuning CBR/ABR is something I have planned to do in 3.100.

SimplePortal 1.0.0 RC1 © 2008-2018