Amazon MP3 - Lame 3.98b ABR 249 kbps, Should this be lowpassed at 16 kHz? Options

[carpman]

To me, it looked like the CD version didn't have much above 16 kHz, so that certainly could be an explanation.

Yeah I wondered about that. Unfortunately I just don't know enough about how LAME works. I guess if there's not much energy in the 16kHz+ range LAME presumably says that stuff will be drowned out and so we can despense with it regardless of the 19500 (if Encspot is right) lowpass.

Purely guessing.

C.

[shadowking]

The transcode quality from a 128k mp3 source is catastrophic. Very loud ringing artifact throughout - so one would immediately notice it. It would sound like a bad shoutcast stream.

The same dillema applies to buying lossless - what stops somebody encoding from a 128k source ?

[kornchild2002]

This also seems to be turning into a thread about looking at graphs rather than listening to the music.

Cheers, Slipstreem.

I don't think so. None of us are claiming that we need the information above 16KHz or that we can ABX this lowpass filter. On my part, I am just curious as to why Lame is applying a lowpass at 16KHz with abr 256 while -V 4 uses a higher lowpass even though abr 256 uses a much higher bitrate range. Again, I have not attempted to ABX the results. I am just curious about why Lame does this with ABR yet lower VBR settings use a higher lowpass. That is all. I would agree with you if we were making audio quality claims but we aren't.

If this is normal behavior, it's more like "You can't hear anything above 16 kHz here, so let's not waste bits on it."
To me, it looked like the CD version didn't have much above 16 kHz, so that certainly could be an explanation.

I know that both tracks, in their lossless version, that I posted have a bunch of information about the 16KHz spectrum so that is not the case. I just think that Lame at abr 256 applies a lowpass at 16KHz.

[carpman]

Well, Nerkenheimer's come up with the goods and that pretty much solves my problem.

But like Kornchild, I'm also

curious as to why Lame is applying a lowpass at 16KHz with abr 256 while -V 4 uses a higher lowpass even though abr 256 uses a much higher bitrate range.

C.

[Slipstreem]

I'm curious too, so I can understand where you guys are coming from. But surely there are two obvious answers...

1/ Encode something yourself with the same settings to see if this is "normal" behaviour (that's now been done).

2/ If in doubt as to a third-party's trustworthiness when lossy encoding, buy the album on CD and encode it yourself.

You don't need to stray far outside HA to see the abortions that are the "ideal settings" recommended for LAME by some people who call themselves "experts" but quite clearly aren't. I always buy the original CD out of a total mistrust in anybody else to do it properly without imposing their own personal twist on what "properly" is.

If I didn't do this then I'd always find myself in the same position as carpman, so I do understand his concern. But the resolution to the problem does seem, to me at least, fairly obvious. Buy the CD in the first place.

Cheers, Slipstreem.

[pdq]

I think the main concern at this point is that lame appears to be claiming to do one thing and in fact is doing something quite different. This would be a bug and needs to be brought to the attention of the developers.

Has anyone tried other versions of lame to see if they do the same thing?

[kornchild2002]

I would be happy to test Lame 3.97 if someone can link me to the exe file. I just upgraded dBpowerAMP to 3.98.2 so I no longer have the 3.97 exe file. So right now it appears to be normal behavior with ABR and Lame 3.98.2 as I have since encoded about 20 tracks worth of material to abr 256 and they all have a lowpass at 16KHz (all tracks also have information in the full spectrum).

[timcupery]

I would be happy to test Lame 3.97 if someone can link me to the exe file. I just upgraded dBpowerAMP to 3.98.2 so I no longer have the 3.97 exe file.

Test away:
http://oldversion.com/program.php?n=lamemp3
This has 3.93 and 3.97. I have others at home, too, and can email you the executables if you'd like more to test with.

[timcupery]

So right now it appears to be normal behavior with ABR and Lame 3.98.2 as I have since encoded about 20 tracks worth of material to abr 256 and they all have a lowpass at 16KHz (all tracks also have information in the full spectrum).

I just messed around with a short clip (just over 1 second) of Tom's Diner, and for this, Lame 3.98 appears to lowpass around 19.5 khz.

* Lame 3.98 and 3.98.2 produce exactly the same audio output. I didn't have any other encoders here to test.
* cbr 256 produces almost exactly the same spectral display (and therefore has the same lowpass behavior) as abr 256. (note that abr xxx and cbr xxx are the same as the old preset abr xxx and preset cbr xxx)

The most interesting thing is that for this clip, lowpass drops a little lower moving from abr/cbr 256 to V0. I found this surprising.
I made an animated GIF file (1.25s lag) showing the lowpass shifting through the original wav, abr 256, cbr 256, V0, V1, V2, V3, V4. The most noticeably striking change is between V2 and V3 (presumably where the -Y switch comes into play). Check it out:



edit: added the original wav image to the animated gif. dumb thing to overlook.

This post has been edited by timcupery: Jan 30 2009, 21:23

[lvqcl]

I would be happy to test Lame 3.97 if someone can link me to the exe file.

http://homepage.ntlworld.com/jfe1205/LAME/
lame-3.90.3.zip / lame3.96.1.zip / lame3.97.zip / lame3.98.zip

[kornchild2002]

Thanks. Here is the spectrum of the raw WAV file. I decided to use the song Kill Tomorrow by Mushroomhead, it is off of their "XIII" album which was released in 2004. I decided to use this song because it contains information above 16KHz and would allow us to effectively see what Lame is doing at 16KHz and above. All Lame files were encoded using abr 256. So here is the first spectrum of the lossless WAV file:


As you can see, the lossless source has not been lowpassed.

Here is the spectrum from Lame 3.93:



It has a lowpass filter at 20KHz which, from my understanding, is a standard feature of Lame.

Here is the spectrum from Lame 3.97:



It has a lowpass at 16KHz

Here is the spectrum from 3.98.2:



As we know, it has a lowpass at 16KHz.

This is pretty interesting as the Lame developers have made it where Lame used a lowpass at 16KHz, when using ABR, somewhere along the line. Lame 3.93 has the standard 20KHz lowpass but all others are at 16KHz. So the whole lowpass at that level appears to be normal behavior. My only question is why? Again, I am not saying that I can ABX the results or anything. I am just curious as to why abr uses a 16KHz lowpass filter with the newer versions of Lame. I am not trying to challenge the decisions of the Lame developers, I would just like to know the motive behind their decision.

Edit: I went ahead and encoded the lossless file to 256kbps CBR using those three versions of Lame. The spectrums were the same. Lame 3.97 and 3.98.2 have that lowpass (which some are now saying is not a lowpass but it appears to be one) at 16KHz while 3.93 doesn't. That explains a lot.

This post has been edited by kornchild2002: Jan 31 2009, 03:53

[carpman]

If I didn't do this then I'd always find myself in the same position as carpman, so I do understand his concern. But the resolution to the problem does seem, to me at least, fairly obvious. Buy the CD in the first place.

I've now bought a total of 2 MP3s and have swiftly come to share your conclusion. It's just a real bummer when you only like one of the songs on a CD. The real solution is for retailers to sell lossless - there aren't many cases this extreme for me (i.e. liking only one song) - in this case I'd have happily paid twice the price for a FLAC download (£1.38) rather than spend £4.20 (for the cheapest 2nd hand CD incl. postage).

C.

[halb27]

AFAIK ABR/CBR have the HF behavior like when using the -Y switch with the VBR mode (defaulted with -V3 and below). It's a feature.
The spectograms look like that, there is no real 16 kHz lowpass, but a machinery letting HF above 16 kHz pass only for good internal reasons.

[[JAZ]]

I belive this is getting a bit too far from what it should be.

Not even the song that originated this thread had a real lowpass at 16Khz, and definitely not the ones that have been posted after that.

I want to remark this: A lowpass is not selective. (Ok, musepack had what was called an adaptive lowpass, but that's a different beast).

Sincerely, what you are seeing is LAME working. It may not be perfect (Remember the several threads commenting on LAME 3.98 misbehaving in some samples, especially when targeting low bitrate).

Let's focus on the -Y switch and its relation with one of the limitations of MP3: band 21 (or 12 for mpeg2) not having a scalefactor ( or the sfb21 problem ).

If you look at the screenshots that all of you have posted, only those parts above 16Khz that have loud content are encoded.



Now, let's focus on what the images really tell, and what, if any, has to be done to LAME.

We see a marked section above 16Khz that is more reduced than the rest. We know that there is where the last scalefactor starts.
We see content above this, just when it is loud, and with some holes.
It is noted also, that this is using a high bitrate.

Gains of the current behaviour:

- Less precision needed for high frequencies allows the rest of them to have more bits.
- More bits allows more shorts blocks (pre-echo control), more left/right (in the event it is needed)...

Losses of the current behaviour:

- Visual (This community doesn't care about that).
- Could cause something similar to the dropout artifact that is heard on lower frequencies.
- The loss in high frequencies in genres like metal or other such genres may be higher than the fidelity gained in lower frequencies.

And of course, there's the possibility that it hides a bug somehow.

Especially strange (for me) is the image posted by kornchild on post 17 . I find strange that the frequencies around 20Khz are (visually) better represented than those around 17Khz.
I wonder, then, if somehow the encoder is not seeing that correctly.


Thus, if we want to get this further, we need more samples like that, and that the developers or people in-the-knows about the psychoacoustic model to point to things to verify.

Because, and this is what remains, currently we don't know if what it does is right. Comparing it with older versions doesn't necesarily give us an answer, because this is about lossy encoding, not graphics.


Edit: I have the feeling that the -Y switch is acting the same way on both ends of the quality (or bitrate) range. Thus, providing too much quality (and too few bits) on low bitrates, and too low quality (and too much bits) on high bitrates. But we have no proof.

This post has been edited by [JAZ]: Jan 30 2009, 23:20

[ameyer17]

Especially strange (for me) is the image posted by kornchild on post 17 . I find strange that the frequencies around 20Khz are (visually) better represented than those around 17Khz.
I wonder, then, if somehow the encoder is not seeing that correctly.

Could it be dithering/noise shaping from the decoder?

[kornchild2002]

Especially strange (for me) is the image posted by kornchild on post 17 . I find strange that the frequencies around 20Khz are (visually) better represented than those around 17Khz.
I wonder, then, if somehow the encoder is not seeing that correctly.

Which image are you talking about? Each image represents a different song, one was purchased from the Amazon mp3 store while I encoded the other one from a lossless file. Both songs are different though in that they come from different artists, albums, etc.