Exactly, I always decide "is it worth it" and choose between lossless and mp3. When using mp3s, I want to use high quality with the good feeling, not to have wasted bits, hence a high quality vbr preset.


This sounds like a good idea. I did some testing, the abr behaviour seems, erm, limited:

--abr 256 -b192



--abr 284 -b192:

-- abr 284 is too much according to my personal taste.
I admit my paranoia is such that I'm quite pleased to see at least some 320 kbps frames except for music I consider very easy to encode.

You can achieve this by using abr x with x>=260.
Use for instance --abr 260 -b192
or --abr 270 -b224.

As I said abr is pretty safe and pretty stupid.
With abr 284 you practically force lame to use that many 320 kbps frames cause lame abr does not want to deviate very much from the bitrate you ask for.
So your abr 256 result looks more realistic as for what is needed. As the average bitrate is below 256 kbps whereas you asked for 256 lame has considered your track being a rather easy one to encode. I don't know much about Bruckner but I guess lame is absolutely right. If you're still sceptical just do abxing. It's easy to do with foobar (but it's a very hard job).

My personal favorite setting with my productive encodings was
--abr 270 -b224 -h --lowpass 18600 (Lame 3.90.3)
and I have never had a problem with that. I did a lot of abxing with normal music and different serious problem samples and erverything has always been fine.

As you can read in the 'transition band and lowpass' thread I will use another lowpass in the future, together with slightly reduced bitrate demands, but that's a rather personal setting.

y'all know the routine, don't tamper with the presets. At least not just to get a nice looking bit distribution graph. I'm pretty confident they don't beat the presets quality wise.

I'll never understand why some people care more about how their music looks like, rather than how it sounds like.

Oh, and - raising minimum bitrates is really cool - especially with tracks which have 1min of silence in them - that was really cool when i encountered it - 1min of high-quality digital silence encoded at 192kbit - just to be on the safe side of course *eyeroll*.

- Lyx

Are you listening to minutes of digital silence on a regular basis?

*yawn*


No.


Really sure? There were lame vbr presets with having a -bn value but which encoded dig. silence at 32 kbps. Apart from that: No, most tracks haven't got digital silence for minutes. Don't tell me about exceptions, I know about all the exceptions I own on CD.


We all know, and you can read the original post is not about tampering presets. But on the other hand I see, if we ALL always state "trust lame", "trust the presets", "it's all good as it is", there would have been less development. What I felt before posting was "the bitrate distribution looks as if theres a quality preset (-V -1) missing". Not more, not less.

This is a very good point.

When I came to HA last year I had this high quality demand in mind which you're out for.
I had some questions especially in the beginning which certainly didn't show up deep knowledge on mp3. But I did learn and meanwhile I'm quite sure I had my personal little contribution for helping improve lame (bit reservoir was used a bit strangely or not at all up to 3.97b1 but was corrected with b2 after the discussion between Gabriel and me; problem sample behavior got into special focus because of contributions from several members and I was one of them and because Gabriel did take care of it and improved it with 3.98a3).
Sure there's never universal wisdom. Many things are a matter of taste (my -b setting together with abr for instance which is not based on any listening experience), and some things are not (lame 3.90.3's superior quality on problem samples like trumpet.flac for instance - valid at least today compared to the currrent lame versions).
And of course targets are different. People who care a lot about wasting bits will use other settings than people who don't but are out for the utmost achievable quality.

Are you trolling on a regular basis?

Of course I - and I assume Lyx as well - don't listen to digital silence, and that's exactly the reason why I don't want LAME to waste any bits on it. Hence, no -b tweaking for me...

just trust LAME devs, they are smart guys ... this was at -V5:

Yes, I'm competely braindead and like watching gaussian bell curves the whole day. Thank you.


Yes.


The chart I posted in the thread opener was to show that there could be a higher quality vbr preset/ that there's a big gap between -V0 and CBR320. Yes, we all know vbr works adaptive and bitrate distribution depends on the material.
Now about the wideness of the bitrate distribution. Why is the distribution in --vbr-new so much wider than in vbr-old?

From the lamedocs:

Objections show basic lack of familiarity with the subject

I can understand halb27 (I am a troll then?) with the -b approach. E.g.:Someone posted a pretty bad problem sample a couple of months ago (3.97b era), at the bad point lame chose remarkably low bitrates. halb27 could have chosen -b to defeat such (yet unknown) problems. Whats the problem, why is there need for a dogma? For the price of bigger files he chose the chance of having fewer problem samples.

I'm not sure I remember exactly what thread you're refering to, but basing a decision on a single test sample does seem pretty silly. Afterall there are test samples that, if taken as representitive of all music can support everything from Full Stereo, Not using Full Stereo, using VBR, not using VBR, etc.

I think you're find a problem where there isn't one. There are samples out there that use very little bitrate, either because they have no high frequency content (and thus don't run into the scale factor issue that bloats bitrate in the VBR presets), because they have no real stereo, they have lots of since, or probably other reasons I'm not familar with. That doesn't mean theres a problem. It could just mean you're not dealing with the scale factor bug as is typical with MP3. Or some other reason. So jumping to the conclusion that this needs a fix is a little premature, which is why people are skeptical.

Squeller

can you please try to use Winmp3Packer after your --abr 284 encoded file?

http://www.hydrogenaudio.org/forums/index....hl=winmp3packer


You might find out that those 320 kbps frames are just filled with padding after all.

As for the very start:
I guess the -V0 encoding is perfect. There is no evidence it's not. The bitrate distribution really doesn't tell about that, only abxing can do. Few or no 320 kbps frames don't tell about encoding quality.
But what can be seen despite of this is Squeller wants to achieve the utmost possible quality as the most important target, and doing the encodings efficiently as the next important goal.

@ ChiGung:
Thank you for pointing out that even minutes of digital silence are efficiently encoded with these high quality settings.

challenge @ Lyx:
I know from many of your posts that you strongly hold up the paradigm of 'if you want near-to-perfect quality: go lossless, with mp3 there will always be problematic samples even with cbr320'.
If you want to be taken serious (especially as you are a long time member considered knowledgable) will you please give me a sample on which
Lame 3.90.3 --abr 270 -b224 -h --lowpass 18600
is abxable? I've been searching for such a sample since last october, I contributed to a lot of threads on HA with this kind of encoding, with a lot of people replying with a different attitude than mine, but nobody ever came up with a sample which showed that the setting given above shows up problems.
Go ahead and show me one, or be a little bit more quiet. Nothing is wrong going lossless of course, but it's not the necessary way to do.

@ kwanbis:
The test shows that Lame 3.97b2 is great at even 128 kbps.
But this result is to be taken in a different way than you can do with test results of a certain hifi equipment for instance.
The great results are valid for the samples that were tested. Sure they will be valid for the vast majority of tracks. But the problem sample issue exists. I'm not talking about subtle artefacts barely audible. Go to Gabriel's problem sample and his lame attack thread and you'll find quite a lot of samples of that kind. Try for instance trumpet.flac or herding_calls.flac with 3.97b2 -V5. I guess many people who have heard these distortions will think about these things in a different way. -V0 improves things, but the artefacts are still abxable rather easily. Moreover certain problem samples fall into certain musical classes. Trumpet- or saxophone(-like) sounds is such a class. When encoding trumpets I don't think probabilty for bad encodings is very low.
robert (not Gabriel as I wrote in another post) improved things with 3.98a3, so luckily Lame development is still improving. But we're not at the end of the road: trumpet.flac for instance is still abxable at -V0.

As for my -b and other settings:
I use it and I don't to hide this fact from other members. I'm well aware that this is a questionable thing as I have no evidence that plain abr is not reliable.
My personal history is that I wanted to use cbr320 in the beginning because I don't have to care about file size (but I do care to use mp3).
Later I found (assisted by proposals of ChiGungs') high bitrate abr will give me essentially the same security as does cbr320. Thanks to ChiGungs' proposals I also use the lowpass which can make encoding a bit easier on some occasion while not having a negative impact on quality as this is the lowpassing of -V2 which I have always been happy with.
After having done a lot of abxing I know Lame 3.90.3 -b192 cbr encodings which are not transparent to me though not very annoying. With Lame 3.90.3 -b224 I do not know any sample which I wouldn't rate at least 'very good' (though not transparent in any case).
So using -b224 together with --abr 270 gives me more security based on my -b224 cbr experience. May be this is unnessecary as --abr 270 might always choose high-enough bitrate. I even guess this is the case, but -b224 gives me more of a definite security, and I don't care about filesize anyway.

Surely that was an unwise challenge


Dood, enough thanks - i dont want to liable if it breaks

Why that? It's true I will be a bit disappointed when the first abxable sample comes up, but it's more important to learn about such a sample. And as for Lyx I would really like him contribute in such a constructive way.

Ok. Your thoughts inspired me, but certainly you're not responsible for what I do with them.

Thank you.
Will you please give me two links, cause there are several castanets samples out there, and I don't know the other sample.

http://ff123.net/samples.html
http://lame.sourceforge.net/gpsycho/quality.html
(but nearly all castanets samples should be ABXable)

Thank you.

Just did abx.

As for castanets:
I know this sample from an earlier abx test a few months ago, and then I was not able to abx it, but I did that with rather poor equipment. Meanwhile I got a decent soundcard and headphone, and may be I'm more experienced now too.
I arrived at 7/10, and as I'm certainly not the champion at abxing (especially pre-echo), sure castanets is abxable with my setting.
On the subjective scale however I rate the deviation from the orginal as extremely subtle and not annnoying.
As the pre-echo problem with castanets is a well-known problem to any transform encoder I rate the result as 'very good' though you're absolutely right giving it as an abxable sample.

As for kraftwerk.flac:
I found the sample after looking at ff123's sample page before I read your answer, hoped it is your sample, and did my abxing instantly.
Usually I have problems concentrating on electronic music (cause I don't like it) but this sample was not annoying to me.
Anyway I was not able to abx.
Gabriel, were you able to abx my setting on this sample? Can you give me a hint?

I was thinking about the kraftwerk sample that had been used in this test:
http://www.rjamorim.com/test/multiformat128/results.html
If the one offered by ff123 is the same, then mp3 should have extreme difficulties with the clicking pulses.

The samples from that tests are not available any more, but there's a good chance it's the same sample cause rjamorin refers to ff123 in the overall context.
As for the sample you referred to by your link I guessed the problem is with the pulses. I went into another test but really can't find any deviation from the original.
Which just says I can't abx this sample with my setting. Anybody out there who can?

Methodologically speaking however I did expect when you're talking about samples being abxable with my setting that you (or somebody you definitely know has done so) did abx my setting on these samples. Obviously this is not the case. You don't even know whether the link you gave me points to the sample which you have in mind. But I don't want to go into this too deeply having made a lot of mistakes myself (starting being a member here with strange settings in mind, continuing with a serious memory lack concerning a previous abx result of mine, and finishing [so far] with placebo feelings accompanying my 3.98a3 trumpet abxing [I could abx trumpet.flac with -V0 --vbr-new as well as plain -V0 but really felt it harder to do so with plain -V0]).

You know pre-echo problems are immanent to transform encoders especially mp3 and just gave me two of them which you remember should be serious.
There is an absolute and a relative relevance to the pre-echo problem.
As for the absolute relevance:
All those who are so sensitive with respect to pre-echo that they are not content with for instance my setting should really go lossless (or try a very high bitrate lossy codec like wavPack lossy).
Everybody can find out for himself with the two samples you gave, Gabriel.
I personally am absolutely happy with the results.
As for the relative relevance:
Do you know a Lame version and setting which behaves better than does my setting on the two samples you gave with your links? Any real abx experience?

Badvilbel is better than vbr but still not quite right even with --preset halb27. Mind you this is one of the most extreme samples ever known.

foo_abx v1.2 report
foobar2000 v0.8.3
2006/02/12 23:25:19

File A: file://H:\temp\abx tests\badvilbel.flac
File B: file://H:\temp\transcoded\badvilbel.mp3

23:25:23 : Test started.
23:26:07 : 00/01 100.0%
23:26:26 : Trial reset.
23:27:09 : 01/01 50.0%
23:27:32 : 02/02 25.0%
23:27:39 : 03/03 12.5%
23:28:05 : 04/04 6.3%
23:28:10 : 05/05 3.1%
23:28:15 : 06/06 1.6%
23:28:21 : 07/07 0.8%
23:28:26 : 08/08 0.4%
23:28:28 : Test finished.

----------
Total: 8/9 (2.0%)


I don't think these custom switches are pushing quality up at all - rather its a 'brute force' when the psymodel is failing. I'd rather gabriel and robert fix the psymodel than rely on these commands.

Thank you for your test. Will do it too.
You did not talk about the lame version you used just about the settings. Did you use 3.90.3?

You're absolutely right, these switches do have a brute force approach in mind.
And if things go well we'll see a lame version with a psymodel and a vbr behavior which makes this brute force stuff unnecesary. It's on its way already with 3.98a3 as the first outcome.

I can easily be misunderstood because I am so strongly proclaiming my setting.
Honestly speaking I have two targets in mind.
First of all I'm not pretending anything. My setting is my productive setting (apart from a rather personal different lowpassing which I will use in the future), and everbody who has a similar utmost-quaility target in mind might find it useful too.
But this is not the ultimate setting. I'm very happy that current Lame development does improve vbr behavior and psymodel usage. All the things I am not content with have been more or less fixed with either Gabriel's lame-attack or robert's 3.98a3 version. So there is a good chance that these things are brought together in a satisfying way in the not-too-far future. I admit I am also trying to be a bit of a sting for this development, but I hope I do so in a more or less balanced way trying to give the valuable work of the lame devs the right worth despite temporary discrepancies.

The wonderful thing is: lame quality is still improving no matter that there is a limit for sure. Look at the recent 128 kbps test, look at what has been achieved with 3.98a3.
Using cbr320 because of problem cases was my starting point but according to my experience a defensive very high bitate abr setting does the job as well. Sure with this I allow for a certain danger that this is not sufficient whereas cbr320 is. For real-life tracks with abr 270 there is a rather high percentage of 320 kbps frames so it's very likely that the problematic situations are covered by them. Sure it will be possible to construct a 5 second problem sample with a pre-echo challenge all of the time, but I don't care about such samples. As for your sample: do you think it improves using cbr320?

Your sample unfortunately belongs to the kind of electronic music I can't concentrate on (at least I got a 6/8 which turned out to be a 6/10 as I usually do 10 trials). Your abx result is more valuable than mine anyway.

I searched a bit for badvilbel and learnt it is a serious problem sample even for codecs like very high bitrate wavPack lossy.
What I could not find is: is the problem also due to pre-echo as is the case with Gabriel's samples? Due to the characteristics of the sample I guess it is. Or are there other reasons for the problem?

You were able to abx your sample with the abr setting. From the way you were talking about it I interpret it as: the deviation was not strong especially measured against what you expected from this sample. Is this correct, or how would you describe it?

And: do you know a lame version and setting which behaves better?

There is no need to guess; it is the same sample. If you go to the above URL and scroll down the page you will see detailed info to the effect.

I failed to read that the sample was submittted by ff123. So it will be the same sample.

Well, halb27 when listening Kraftwerk and Lame 3.90.3 --abr 270 -b224 -h --lowpass 18600 there is a very BIG artifact at second 8 and 13! Lame 3.93a is much better even with V2 there. How does it come you don´t hear this? When you want to recommend something over here you should do better.

Edit: typos

General use of 3.98a3 is not yet recommended because its an alpha, using it as a benchmark is ridiculous considering its a fresh breakthrough
How does the artifact in the sample compare to other recommended versions and presets? You'd do better to be less impatient to score a point.

I am not on a mission finding the new best setting with old encoders and won´t test in this direction, sorry.

Test it against the current recommend version then, and if you wouldnt like to come across as unreasonable, if it doesnt fare badly, a retraction of your previous criticism which misapplied the performance of the latest breakthough alpha -would be wise.

Or maybe you just want to leave it because you suspect your criticism of Halb and his command line wont stand up to a fair test?

Thank you, Wombat, for your contribution.

Thank you, ChiGung for your support, but as Wombat has found a problem with my setting whereas 3.98a3 (@Wombat: is this correct? you were writing about 3.93a) performs better this is valuable information.
Moreover as I know from other tests his listening abilities are much better than mine so he's very much welcome to contribute like this.
I will try again tonight.

Let me say it like this: in case samples come up where the new lame versions behave much better than my setting wouldn't that be great? Best would be samples that are not abxable for sure. In case superior behavior is with 3.98a3 it would mean it's not very wise to use it right now but we can look forward to a productive version with this great quality.

What I do is: I take my setting as a working hypothesis that it's the best setting so far for productive purposes. (best according to: quality is first target, coding efficiency is second, no two phase procedure like shadowking's preference using V2 as a first phase, and in case it fails, cbr320 as the second).
I have no problem (ok: not a big problem) being proven wrong. And the perspective of getting a better productive version is valuable as well.

To be more correct i used a newer alpha, 398a3-3 that could perform a little worse than plain 398a3 cause the bitrate was brought down again to match 3.97b2 bitrate.

@chigung
halb27 asked about samples that hurt his abr command line and was told to test kraftwerk. I just verified it out of curiousity. If you want to have more results with different encoders, just do it.

halb27, to get a perspective of what your settings can (and can't) do, try it with the clips dibrom used to tune the presets:


Maybe you will find that the presets indeed solved more problems than they created; it is always good to test many different problem samples, not just focus on a few isolated cases.

As you are talking about a very BIG artifact I'm really upset having abxed right now without success since quite a while second 7.0...9.2 and 12.1...13.9.
Are we really talking about the same sample? Can you please verifiy whether my sample/encoding is what you were listening to?:
kraftwerk.mp3 (temporary upload)
Thank you for your efforts.

Yes, it is absolutely the same file. Second 8 should be easier than 13. It may help you to listen at something like digital errors when reading scratched CDs, not like clicks but ugly distortions. I think it is big. But that may differ. This is no pre-echo it is pure added noise.

I confirm Wombat's claims. For me the problem sounds like strong smearing (there's relatively big noise packets before the attacks). If you can't hear it, it would simply confirm what Gabriel said about your insensitivity for pre-echo/smearing issues, common with MP3, whatever the LAME version, the command line or the bitrate (as I said, ABXing castanets even with freeformat at 640 kbps isn't a big challenge for a trained person).

Well, I really don't hear it despite the detailed hints (thanks, Wombat and guru).
There's a big noisefloor all over the track, but I can't hear a difference between the original and the encoding.

I don't remember Gabriel having said so, but for the fact it is correct: obviously I'm not sensitive to pre-echo/smearing. After having been on the edge of being able to abx castanets (which was a bit of a surprise to me as I did before without any success) I'm really disappointed especially as the artefacts on this sample are said to be big.

Thanks, sTisTi, for your suggestion on other samples dibrom mentioned. I downloaded those I could get, and found it's all electronic samples of the kraftwerk kind (very roughly speaking), but esthetically much more problematic to me. After my kraftwerk experience I think it's no use to me trying them as the pre-echo issue is the major problem behind them according to the descriptions.

Well, I asked for samples on which my setting could be abxed.
Gabriel contributed with two such samples, and shadowking with another one.
If I classify them correctly these are samples which are expected to be abxable with the result of any mp3 encoder.
Anyway they are valid answers to my question.

With samples abxable with any mp3 encoding the natural question arises (which I didn't ask for when I tried to challenge Lyx): does my setting produce a good result?
Which is more or less equivalent to asking: is there a better way of doing it?
While this an interesting question I cannot expect that somebody does the effort and finds out for me. I personally obviously am not able to answer it for these samples.
And I thank you very much, Wombat, as you gave an answer for kraftwerk (in favor of current 3.98 development).
Anyway, if somebody likes to contribute on this question this is of course very much welcome.
As are samples not abxable with any mp3 encoder result but with my setting.

@halb27: Fatboy is an easy ABX (16/16) with your proposed settings/version. The resulting file consists almost entirely of 320k frames. It really is an extreme killer sample but matches your request.

http://lame.sourceforge.net/download/samples/fatboy.wv

What do you expect? That old members will spend their free time in order to prove that an eccentric commandline isn't completely transparent? It has been established by well-known members that during the --alt-preset maturation (i.e. 3.90 era) that --preset xxx (ABR) were inferior to VBR presets. It's not to Lyx or other people to prove to yourself that you're wrong. Samples ABXable with your MP3 setting are countless. Because samples ABXable even with CBR 320 are countless. MP3 has native issues. It's perfectly true that 3.97 beta introduced new problems, but it also bring other improvements in the different quality area. You're basing your opinion on the few samples you're able to ABX. That's praiseworthy, understandable and fully objective. But you must understand that your own quest of the "perfect MP3 setting" makes few sense for many people here who understood that whatever the encoder, whatever the setting, MP3 encodings can't reach full transparency. That's why there's no reason to make a dramma on samples like trumpet.wav: there are tons of cousins of castanets.wav that MP3 won't ever handle. If you're looking for transparency with lossy encoders, you may have a chance with AAC, MPC, Vorbis or more certainly with DualStream/WavPack lossy, but MP3 is definitely doomed to an eternal failure.

And what is the most ridiculous: looking four years in the past to find a solution to a fresh bug or testing the latest available build? Recommending 3.90.3 --abr 270 -b224 -h --lowpass 18600 or 3.98 alpha with -V2? Sitting on a dead branch or cultivating the new one?
The recommendation is only here to help people who aren't interested to perform themselves listening tests and to guide them toward the right direction. In that sense, using an alpha for reference makes much more sense than using 3.90.3 ABR with a foolish high bitrate floor which hasn't been recommended by someone here in more than four years.
And for your own knowledge, you may be interested to learn that the first recommended LAME version of HA.org was precisely... an alpha version of 3.90. And that people like halb27 have been quickly banned from this forum for having insisted on unwise and outdated (like --r3mix) settings instead of the recommended ones. Times have changed, and now people apparently can recommand what they want without any solid basis and without being annoyed

Can someone explain how Lame 3.90.3 --abr 270 -b224 -h --lowpass 18600 could offer better quality than Lame 3.90.3 -b320? Why would such a (very) high ABR setting be preferable to 320 Kbps CBR? Is it because 320 Kbps CBR handles Joint-Stereo differently to all VBR modes?

IIRC, according to halb27 the modified ABR commandline produces the same output quality as CBR 320 but with less bitrate. An interesting question would be to know:
- on what sample(s) -b192 produce a lower quality than -b224
- on what sample -h is improving the quality
- on what sample is based the 18600 Khz lowpass value.

As far as I can remember, presets were built in order to prevent encoding alchemy (and TOS#8 to request from members to bring objective proofs to their claims). There's nothing wrong with custom commandline. It could be a source of new improvements. But at least we should see concrete elements of experimentation to take seriously them into account. Were they posted? I have apparently missed the good topic.

Just one thing I notice...

I also didn't hear kraftwerk's artifacts encoded by LAME 3.90.3.
So at first I'm confused reading this thread.

After trial and error, I finally find the cause of this issue.
It is inaudible with SENNHEISER HD580 but really audible with Sharp HP-MD33.
Once I heard smearing of this sample, I became to hear artifacts even with HD580.
But it's very minor compared with HP-MD33.

The person who can't distinguish artifacts should change your equipment in some case.

Edit: typo

I know about this issue and used alessandro ms-2 as well as sennheiser px200.
Also listening volume can have a serious impact to me.
However despite all these trials I really couldn't hear it.
But you inspire me to something else. I also made the experiencce that once I heard what is wrong I can hear it also under circumstances where I couldn't before.Will go into listening to a low bitrate encoding as a warm-up.

Thanks for the hint.

You could always look for more samples that show severe pre-echo issues (perhaps testing with another encoder that has more problems than LAME) and see if you can thear it on any of them. Once/if you can do that, it makes it a lot easier to identify on other samples.

It doesn't produce better quality of course. 3.90 --abr 270 uses the same stereo mode as does -b 320, but another one than --alt-preset insane. (The --alt-preset question is the hardest question to me. abr 224 produced better results than --alt-preset 224, but it's true: it's based on a few samples only those with the kind of distortions I consider worst).
It's as guru said: --abr 270 just produces smaller filesizes (some 15% smaller).
This is still much more than is usually needed (as guru complains), but yields 320 kbps frames where they are needed with a very high degree of security (see for instance Gecko's fatboy result: 84% 320 kbps frames, 307 kbps average bitrate. Thank you Gecko for your contribution.).
In a sense abr is something like an unintelligent but rather safe vbr mode.
The "real" vbr mode has a weakness as it cannot really know what is good-enough, but current 3.98 development is about to improve things a lot and find a good compromise towards defensive and efficient behavior.