Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: LAME problem samples - discussion (Read 46346 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

LAME problem samples - discussion

Discussion regarding problematic samples from this thread should take place here.

LAME problem samples - discussion

Reply #1
Forgive me my ignorance, but when is a sample to be considered problematic? When it's not transparent at V2 or better settings? Or...
"We cannot win against obsession. They care, we don't. They win."

LAME problem samples - discussion

Reply #2
When its quality is unusually lower than the overall quality reached by other samples.
It might be something easy to ABX at -V2, but could also be a strong unusual distortion at 128kbps, as an example.

LAME problem samples - discussion

Reply #3
One pretty interesting sample is missing in the Lame problem samples thread.
Guruboolez offered it in this thread:
http://www.hydrogenaudio.org/forums/index....opic=39161&st=0
It is very easy to hear and messes up 3.90.3 aps 397b -V2 and 397b -V2 --vbr-new.
3.96 -V2 is not better also. But 3.96  -V2 --vbr-new does a really good job with it.
Maybe there can be found a relation between 396 vbrnew and 3.97 vbrnew behaving different.
It is a sample from guruboolez and not uploaded here afaik.
So if you can read this guruboolez may you add this to the Lame sample thread somehow? So more people can find and hear it.

Edit: lol! Forgot the name of it  It is S53_wind_saxophone_a
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #4
I posted a link to the "herding_calls" sample. In my test it produced an audible artifact with 3.97b1 and 3.90.3 at -V2 and -V0. The artifact was more pronounced with --vbr-new.

Also, it produced a different artifact with LAME CBR presets below 192.

I found the sample transparent with CBR 192 kbps and higher, but not with any VBR setting I tried.

The previous discussion and test reports are in this thread: http://www.hydrogenaudio.org/forums/index....showtopic=37003

LAME problem samples - discussion

Reply #5
Pretty interesting. Furst time i hear it.
None of the 396/397 V2 settings or 3.90.3 aps can handle it but 396 -V2 --vbr-new is  the best to me.
Pretty similar to the wind_saxophone sample.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #6
Out of curiosity i encoded some files that have the same style like Birds. 
A constant singing female voice in a recording that isn´t absolutely clear mastered.
Elend does such music. On the second album i tried i found it again with 3.97b2 -V2 --vbr-new
This sandpaper noise with birds is for sure not only an exception.
The short sample can be found here.
Deploration
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #7
Hello again here is another sample adding the noise i often hear with 3.97 -V2 --vbr-new
I tried some few music pieces with wind instrument inspired by the wind_saxophone sample of guruboolez.
Et voila! This piece of song is about 2 minutes long and adds it often. As meanwhile common 3.96 -V2 --vbr-new is pretty clean! Here a small clip:

Moon

As this is called a discussion thread i am just beginning to wonder myself a bit. Does nobody hear any problems in Wind_Saxophone, Deploration, Moon and maybe Birds? Maybe someone even did compare these samples to older lame versions? Deploration and Moon was found encoding exactly 3 average Albums with sounds i thought would make problems. Is my auditory canal bent?
Since these problems are pretty easy to repeat it is not a typical "killer sample" problem anymore to me.
As this is not happening near as much with 3.96 -V2 --vbr-new i´d like to hear someone elses mind.
Anyhow this is in strong relation to the next Hydrogenaudio recommended lame version people trust in since years!
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #8
Quote
As this is called a discussion thread i am just beginning to wonder myself a bit. Does nobody hear any problems in Wind_Saxophone, Deploration, Moon and maybe Birds?
[a href="index.php?act=findpost&pid=348712"][{POST_SNAPBACK}][/a]

I can't hear anything wrong with any of these samples. Except that the original is a bit noisy by itself... What should I be listening to? You say sandpaper... to me that means low to medium frequency noise. Is that correct?

LAME problem samples - discussion

Reply #9
Quote
Quote
As this is called a discussion thread i am just beginning to wonder myself a bit. Does nobody hear any problems in Wind_Saxophone, Deploration, Moon and maybe Birds?
[a href="index.php?act=findpost&pid=348712"][{POST_SNAPBACK}][/a]

I can't hear anything wrong with any of these samples. Except that the original is a bit noisy by itself... What should I be listening to? You say sandpaper... to me that means low to medium frequency noise. Is that correct?
[a href="index.php?act=findpost&pid=348880"][{POST_SNAPBACK}][/a]

Sandpaper, maybe is not the best description. It is an added noise like distortion. If you still have an analog telephone this noise is common.
On wind_saxophone it is very constant cause of the constant note the instrument plays in the end.
On the other samples it is pulsating with the music. Very constant in Moon_short. I would rate it into the mid frequency spectrum.
Maybe you have an equalizer set somehow to amplify the highs and it is more masked this way?
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #10
Maybe I have bad hearing, or I am very tired. However I cannot hear anything wrong from the samples.

Would you please give me some hints on what and when to focus on?

Thank you.

LAME problem samples - discussion

Reply #11
@Wombat:
I just listened to Birds, Deplorations and Moon encoded with 3.97b2 -V2 --vbr-new.
I can hear the distortions rather easily.

You seem to be a bit sad because not many people take care of these samples. I consider this thread a good idea of Gabriel's to collect unusually bad encoded samples so that Lame development can improve. For that purpose it's not necessary that many people confirm your results. I don't think anybody doubts what you hear.
And if there are people who can't hear the problems maybe it's because of equipment. As for that I have found it makes a big difference which headphones are used on a specific sample. I usually prefer my Sennheiser PX200 for abxing, but with your samples it was easier to me using Alessandro's MS-2.
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #12
After quite a lot of tests with problem samples I don't think any more I have to use 320kbps for good quality which I beleived I had to do during most of the last months.

So I tried to find out problem sample behavior with lower bitrates.

I used
- trumpet
- herding_calls
- atem-lied
which are the most problematic samples to me.

I used Lame 3.90/3.91 (actually 3.91) which according to my experience is expected to give the best results at the moment (but Gabriel is about to improve things a lot with current Lame development).

I used 'production parameters' which to me means lowpassing to 18600 Hz and to use the -h switch when using --abr x or -b x.

So I tested for x=192, 224 and 256
--alt-preset x --lowpass 18600
--abr x -h --lowpass 18600
-b x -h --lowpass 18600.

Results concerning the different usage of a certain bit rate:

--alt-preset 192 was clearly inferior to --abr 192. At 224 kbps the difference between the two abr modes was smaller, and at 256 kbps it was more or less neglegible.

I couldn't really tell -b x quality apart from --abr x, it was only in one case (herding_calls at 192 kbps) that I had the impression --abr x was a little bit harder to abx.

(As a sidenote: I couldn't resist testing 3.91 aps and ape on the samples. Results were as horrible as expected.)

Results concerning the different bit rates:

192 kbps:
Results were not transparent as was expected, but abxing was already not easy for me (with --abr 192 and -b 192).

224 kbps:
Not transparent either, but abxing was already difficult to very difficult.

256 kbps:
At least at the border to transparency to me. Exactly speaking: Within the effort I was willing to apply within this test I could not reliably abx. There were trials where I was pretty sure I was guessing right. So may be things change when spending a lot more effort. However I'm not sporting for abx mastership - to me there's no big difference between 'not abxable' and 'abxable but only with extreme pain'.

Taking it all together:

For practical listening situations where concentration is a lot lower than when abxing
3.91 --abr x yields very satisfying results for x >= 192, and quality scales well with x. For x >= 256 quality is more or less perfect in a practical sense (as long as the samples considered are really something like the worst samples).
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #13
@Wombat:
Do you also hear the sandpaper problem with -V2 ? (not vbr-new, just old vbr)

LAME problem samples - discussion

Reply #14
On most samples it is also there with plain V2 but always to a less degree. I don´t have the single samples on my mind and i am not at home to test it more specific.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #15
Rarewares offers 3.98a3 now and the lame developers spent much work regarding these problem samples it seems.
I can´t abx birds, moon_short and deploration anymore, but i have to add that i am very tired.
herding_calls, wind_saxophone and the trumpet sample became MUCH better.

After all this is very promising and shows the devs know exactly what they´re doing

Edit: I am  talking about v2 vbrnew
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

LAME problem samples - discussion

Reply #16
Good news!

Will try tonight.
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #17
Well, looks like this version is really a step ahead towards extremely good reliability.

Wombat has no problems any more with 'his' samples, and as he said already trumpet and herding_calls have become considerably better compared to former 3.97/3.98 versions. I can confirm this.

Having heard and encspotted this version I can also understand Gabriel saying VBR is not to blame. With this version it is true for trumpet and herding_calls: very high bitrate abr modes don't behave better than -V0, with trumpet VBR is aware that it has to go high with bitrate (I assume the same for herding_calls but cannot conclude this from the Encspot information).

Gabriel and whoever else has contributed have done a lot of effort in improving Lame's behavior, and they have done so with good success.

So everything's fine now? Not entirely.
At least not for trumpet or saxophone like sounds of a certain kind.
trumpet isn't very annnoying any more with v2 vbrnew, but it's still easily abxable (at a resulting bitrate of 248 kbps). Quality improves with v0 vbrnew (274 kbps), but abxing isn't very hard to do. v0 (vbrold) gives a little bit better quality to me than v0 vbrnew but it's no problem to abx.
herding_call's results are a bit better to me than trumpet's.
3.90.3 high bitrate --abr x behavior is definitely better - perfect to me for 256 kbps.

@Gabriel:

As you said: looks like the psymodel is the limiting issue.
I guess nspsytunes has a lot of advantages in the low to moderately high bitrate range, cause otherwise you woudn't use it. But why not use gpscho in the very high bitrate range, with VBR the 3.98a3 way or similar?
High bitrate might compensate for the one or other disadvantage gpsycho may have compared to nspsytune. Other than that gpsycho seems to be very good at pre-echo and attack problems which are major issues with the usual perceptual encoding techniques. Why give it away totally?
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #18
Quote
As you said: looks like the psymodel is the limiting issue.
I guess nspsytunes has a lot of advantages in the low to moderately high bitrate range, cause otherwise you woudn't use it. But why not use gpscho in the very high bitrate range, with VBR the 3.98a3 way or similar?

Developping two psychoacoustic models at once requires VERY intense work -- this isn't really feasible on spare time ...

LAME problem samples - discussion

Reply #19
Quote
Developping two psychoacoustic models at once requires VERY intense work -- this isn't really feasible on spare time ...

Sure, and it's easy to say so having not to do it.
Of course everything is left to the developers according to their time and according to whatever they think is right.
Anyway luckily they're still about to improve Lame. I'm sure my postings have not always been welcome, but they might also have contributed a tiny bit in thinking a bit different the one or other way. That's how progress works (and I can see this with my own job. There are times when I can't continue well my with my work [database development]. I'm used to talk with a certain colleague then. He's not an expert, but he's really in the game and asks me the right questions to make me see things in another light).
As for gpsycho development maybe there is not to be done extremely much as there is a good basis for that with Lame development up to 3.91. May be results are great already with porting 3.98a3's VBR behavior to the gpsycho implementation of 3.90.3.
Exactly speaking: Why not use two strictly different encoder technologies within one encoder of the kind: for -Vx up to -V2 use 3.98a3, for -V1 and -V0 use a VBR-improved version of 3.90.3. Things can merge together in case this should be considered good over time.
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #20
Quote
Exactly speaking: Why not use two strictly different encoder technologies within one encoder of the kind: for -Vx up to -V2 use 3.98a3, for -V1 and -V0 use a VBR-improved version of 3.90.3. Things can merge together in case this should be considered good over time.

Rather than merely "switching between 2 versions", it would be much better if the goods of the 2 are combined into a single piece of code. This is development. Now the "culprit" is found, developers should focus on fixing it and improving it, instead of spending their already limited resources on other merging work.

LAME problem samples - discussion

Reply #21
Quote
Rather than merely "switching between 2 versions", it would be much better if the goods of the 2 are combined into a single piece of code. This is development. Now the "culprit" is found, developers should focus on fixing it and improving it ...

The progress of recent years has been achieved within a certain framework, with nspsytune as a vital element.
So I guess it's not a short term option (and propably will never be an option) to generally use gpsycho.
The progress reached must be kept.
But development AFAIK was focussed on VBR in the moderate bit rate range.
-V1 and -V0 have never been attractive options, as ususally -V2 or lower was very good, and in case it wasn't -V1 or -V0 didn't change things essentially (roughly speaking).
So nothing is really given away in terms of 'progress reached' when giving -V1 and -V0 a different meaning and use gpsycho as a basis for these options.

As for gpsycho usage with current state of development there is a problem.
With the alpha version access to gpsycho is available. I did a short test with gpsycho and -V2 last night because 3.98a3's VBR is very promising. The result according this short test was very good.
But of course I will not use this combination for productive purposes. I even hesitate to continue the test. The reason is that as long as I don't know more details about gpsycho usage within 3.98 it's like playing roulette. All I know is gpsycho development has been discontinued, so the overall encoder behavior is undefined to me.
Sure the devs may know better, but I don't know anything about whether or not it can be considered reliable to use gpsycho within 3.98a3. I don't like playing roulette.
Because of other focus may be the devs are in a similar situation.
But as gpsycho development had been vital up to 3.90, it is an option for -V1 and -V0 to use exactly the 3.90 way of doing it (using exactly that source code), and in a first step to improve only the most obvious disadvantages (VBR behavior IMO) according to actual knowledge.
More improvements can emerge over time.
lame3995o -Q1.7 --lowpass 17

LAME problem samples - discussion

Reply #22
I downloaded Lame 3.98a3 from rarewares and I did some ABX tests with him.

I tested the infamous aps_killer_sample, which caused serious and annoying artifacts with Lame 3.97b2 and 3.98a2 with -V2 and -V2 --vbr-new. Even without ABX, this was noticeable.

For my surprise, the annoying problems with this sample seem to have disappeared with 3.98a3.

I obtained with 3.98a3[-V2] a score of 4/10; which is not significant; transparent for me.

Very interesting results in bitrate:
aps_killer_sample:

3.97b2[-V2] --> 182kbps (very annoying artifacts)
3.98a2[-V2] --> 182kbps (very annoying artifacts)
3.98a3[-V2] --> 205kbps (transparent)

3.97b2[-V2 --vbr-new] --> 181kbps (very annoying artifacts)
3.98a2[-V2 --vbr-new] --> 181kbps (very annoying artifacts)
3.98a3[-V2 --vbr-new] --> 205kbps (transparent)

NOTE: [-V2] in 3.98a3 is now --vbr-new; in other words, -V2 and -V2 --vbr-new are now the same; but this --vbr-new has been improved and now he already doesn't suffer of bitrate drop that caused problems with some samples... Excellent work!!

LAME problem samples - discussion

Reply #23
Harp40_1

This old reference sample produces the worst LAME artifact I have ever heard at -b 128. The first two notes are totally destroyed. In Ivan's Nero AAC test I found the 48 kbps AAC encodings much better.

The problem is still clearly audible at - b 160 and -V3 --vbr-new.

LAME 3.97b2 @ -b 160
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2006/02/06 09:08:26

File A: file://E:\test\LAME problem\harp40_1\harp40_1.wav
File B: file://E:\test\LAME problem\harp40_1\harp40_1.mp3

09:08:28 : Test started.
09:09:25 : 01/01  50.0%
09:10:43 : 02/02  25.0%
09:11:06 : 03/03  12.5%
09:11:14 : 04/04  6.3%
09:11:38 : 05/05  3.1%
09:11:46 : 06/06  1.6%
09:12:12 : 07/07  0.8%
09:12:31 : 08/08  0.4%
09:12:39 : 09/09  0.2%
09:12:50 : 10/10  0.1%
09:12:58 : Test finished.

----------
Total: 10/10 (0.1%)


LAME 3.97b @ -V3 --vbr-new
Code: [Select]
foo_abx v1.2 report
foobar2000 v0.8.3
2006/02/06 09:25:25

File A: file://E:\test\LAME problem\harp40_1\harp40_1.wav
File B: file://E:\test\LAME problem\harp40_1\harp40_1.mp3

09:25:27 : Test started.
09:25:35 : 01/01  50.0%
09:25:40 : 02/02  25.0%
09:25:56 : 03/03  12.5%
09:26:01 : 04/04  6.3%
09:26:05 : 05/05  3.1%
09:26:23 : 06/06  1.6%
09:26:31 : 07/07  0.8%
09:26:38 : 08/08  0.4%
09:26:54 : 09/09  0.2%
09:27:03 : 10/10  0.1%
09:27:04 : Test finished.

----------
Total: 10/10 (0.1%)


The sample is available for example here: ftp://ftp.tnt.uni-hannover.de/pub/MPEG/au...am/harp40_1.wav

LAME problem samples - discussion

Reply #24
Quote
This old reference sample produces the worst LAME artifact I have ever heard at -b 128. The first two notes are totally destroyed. In Ivan's Nero AAC test I found the 48 kbps AAC encodings much better.


harp40_1 is an old killer sample, and it is strange that nobody picked it up until now - good that you found it again, as I think most encoders choke on it - I hope it will be helpful to the LAME team.