Are there any published results from ABX tests seeking to determine the audibility of various lowpass frequencies? LAME --aps uses a lowpass of about 19KHz and MPC --standard uses a lowpass of 20KHz. These seem very high to me. I've performed some very casual tests and found that sometimes I can't tell the difference between an original 22.05KHz wav and a 14KHz lowpassed one. Of course, on some music the lowpass needs to be set higher to be transparent.

Why are these lowpasses set so high? Is it because people want to encode the 19KHz signal given off by television sets? I add the --bw 16500 switch whenever I use MPC standard and -Y whenever I use LAME --aps. Sometimes the bitrate reduction from using a lower lowpass exceeds 10%. That's considerable savings and modifying the lowpass doesn't introduce artifacts (though it obviously can reduce transparency).

Can someone suggest some samples I could ABX to see what lowpass I could get away with?

well, probably the first thing to do is make yourself some test signals and see how high you can actually hear. If you can't hear above 18KHz, then 18KHz low pass is a good starting point for you. As I understand what's been said around here, people usually have a hard time telling the difference between a full range signal and one lowpassed a little below their high frequency hearing ceiling, hence the adoption of the 19KHz low pass in --ap-s. So, I would say that if you in particular can't hear above xKHz, then xKHz should be where you start experimenting with the low pass filter.

I got reprimanded by Frank Klemm when I stated that I can't hear above 17KHz. Everybody can hear 22KHz signals, it's just that some people's thresholds of hearing are higher than others.

But I understand what you mean. If you play a frequency sweep and set the volume control so that the loudest portion (somewhere in the 3-5KHz range) is barely tolerable, then you can guessestimate "how high you can hear". For me, the perceivable roll-off occurs between 16-17KHz.

However, can I get away with an even lower lowpass than 17KHz? Just because I can hear a 17KHz signal at a sensible volume doesn't mean I can hear it mixed it with music and masked by other frequencies. Yet LAME and MPC have been highly tuned so the lowpass values used must have come from someplace. I mean, overriding MPC standard's 20.0KHz lowpass with a 16.5KHz one sounds very aggressive, doesn't it?

well, my hearing drops off somewhere between 18 and 19KHz no matter how loud I turn it up. I made some test signals in sound forge that contained only one frequency bursted for one second, alternating with one second of silence, just to see where my hearing petered out. I could hear the 18KHz file ok, but I couldn't hear 18.5 very well and 19KHz just wasn't there for me no matter how loud I turned it up, and the same goes for 19.5, 20, 20.5... Way too many rock/metal concerts when I was a teenager and in my early twenties, damn Motley Crue left my ears ringing for three days, haha. I left --ap-s's low pass alone though, just cuz, heh.

I would try applaud.wav.

bAdDuDeX said that mp3's 16 kHz limitation was very apparent to him on this sample.

ff123

Have you tested yourself on the mustang clips from ff123's audio samples page? (if not, here's the page where he talks about it.)

(edit: oh, here's the man right now. ff123, is mustang still a good test for detecting lowpass?)

I could hear the lowpass on the 14khz-lowpassed clip, I don't think I was able to ABX the 15khz clip, so I'm fairly sure I'll never hear the difference between --aps and --aps -Y.

Does the ~10% bitrate savings also apply to MPC files? I thought the main reason why MP3's saved so much bitrate when they were lowpassed was because of the missing scalefactor (or what it the missing scalefactor band...?) for >16khz frequencies. I don't think MPC has that problem, but maybe taking a huge cut out of the high frequencies still saves a significant number of bits. (I'd run a few tests myself, but right now I don't have a computer with a CD-ROM drive or uncompressed audio on it.)

This is correct, but there's more to it. MPC also uses an adaptive lowpass, depending on the signal energy correlated to the ATH.

Let me illustrate this:
[img]ec0173c51d[/img]

(from "Sting - Island Of Souls", encoded with MPC 1.1, --quality 5)

I highlighted it to make the adaptive lowpassing more obvious. See, there should be no need to manually adjust the lowpass, as it rarely encodes up to the 20 KHz limit anyway.

I've done the same thing and I've obtained exactly same results that you. I can't hear a 19 KHz tone no matter how loud I play it. However, I've not been to many loud rock music concerts, I think this results are pretty common, and would be quite worse for older people.

Using mppenc 1.1 (Have only one CD around - DIO - Killing The Dragon; the wav has content up to 20khz - didn't say that it was useful content )

mppenc --bw 12500 Dio - 05 - Rock & Roll.wav
6,869,672
winamp reports telephone (sounds like crap)

mppenc --bw 14500 Dio - 05 - Rock & Roll.wav
7,638,048
winamp plugin reports preset thumb

mppenc --bw 16500 Dio - 05 - Rock & Roll.wav
8,087,668 bytes
winamp plugin reports preset radio

mppenc Dio - 05 - Rock & Roll.wav (no forced lowpass)
8,420,492 bytes

As far as I know, latest mpc encoders don't rely as much on Filburt's ATH ( http://www.ff123.net/audiogram.html - scroll to bottom :eek: ) when it comes to encoding the HF content, so this might explain the relative small difference between the lowpasses...

So I'd say leave it the way it is and don't mess with the presets

edit: URL

I've come to the conclusion that the maximum bandwidth value isn't something to worry about anymore. If you want to reduce the encoding of extreme (inaudible) high frequencies, use one of the newer ATH curves with a steeper HF roll-off than the default. For instance, --ltq 30 boosts the curve by 30dB at 20KHz. This way, you don't have to establish a hard-coded maximum bandwidth...the ATH will dictate that 20KHz signals are inaudible during musical material anyway. However, if you are encoding a strong 20KHz sine wave, even --ltq 30 may encode the signal whereas --bw 16500 will always chop it off. So the newer ATH curves make more sense.

You can try these samples :

Cosmic Baby - Heaven's Tears - Funny how time flies remix by Jam el mar, taken from Trancemaster 4 compilation.

http://pageperso.aol.fr/lyonpio2001/sample...ass/cosmic.flac (524 ko)

[img]747a09ef4d[/img]

A nightmare for speakers and also for analog tapes. I've discoverded that it's quite impossible to record on a cassette !

Transwave - Vision Quest, taken from Phototropic CD.

http://pageperso.aol.fr/lyonpio2001/sample.../transwave.flac (991 ko)

[img]747a09ef4d[/img]


Special bonus : Highpass test !

Lustmord - Dog star descends, taken from The place where the black stars hang CD.

http://pageperso.aol.fr/lyonpio2001/sample...s/lustmord.flac (207 ko)

[img]747a09ef4d[/img]

All three samples can be looped in order to be listened continuously. Because of MP3 behaviour, loop them before encoding them.

Wow !
Lame encodes the 14 Hz sub-bass in the Lustmord sample :eek:

I can't decode the samples

--
GCP

OK, I replaced the Lpac files by Flac ones, and updated the links above.
If it doesn't work either, I can do Monkey audios too.

Lpac files were fine on my computer (decoded with WA diskwriter).

Works now, thanks.

--
GCP

If someone else wants the samples, download them quickly !

I've got 20 MB of webspace only, and I'm currently adding other samples for an old thread in George Massenbourg professional forum, so it won't be long until I delete these ones.
No problem for me if anyone want to host them in a sample page forever.

That thread was about what songs had their sound killed by digital. I'm going to post Depeche mode- Strangelove, both from Cd and vinyl. I find that the sound of the CD (from Virgin FR) stands quite well for the famous dull and lifeless digital stereotype, compared to a digital copy of the vinyl.
I'll post a thread here with links when it's done.