Hi there,

I did a blind test to a friend which pretended that MP3 was no good for his backup, according to the following scenario:
1. I ripped a song (from Animals, Pink Floyd)
2. I encoded it at 160, 192 and 256kbps with razorlame and the latest lame at the time (summer 2004). I used the r3mix.net template that generate CBR 256kbps files (-q 0 lowpass @19.5, etc...).
3. Burned original along with 3 mp3s onto a CD-Audio. I used razorlame to decode the mp3s to WAV beforehand.
4. Brought my friend the CD
5. He put it in his CD player and turned everything on. We sat on his sofa together.
6. He listened to the first 10 seconds of the song which isn't even music! just random typical PF noise
7. He identified the files in the proper order.

I was stunned.
What mistake did I do? Is MP3@256 still discernable for some people/equipment?

From there I decided to back up my CDs at CBR 320kbps with a lowpass filter at 21KHz and a "-q 0" for all parameters...

Please take a look at the recommended settings for lame.

how many tests did you make? There should be at least 12 samples for each track to make sure he isn't guessing. Also, CBR files aren't very good. Just use LAME 3.90.3 or LAME 3.96.1 with --alt preset standard - no other additional switch since the presets already offer best quality...I doubt he will spot the mp3's using that setting...the bitrates will be around ~200kbps, but still better than some CBR 256kbps file...

It's not necessary trolling.

Pieroxy, you must be careful when you perform a listening test. There are some recommandations to follow. Pio2001 wrote an excellent article about this:

http://www.hydrogenaudio.org/forums/index....howtopic=16295&

Thanks Benjamin Lebsanft, but for the --preset insane setting, I just read that it imposes a lowpass filter at 19.5KHz. If high frequencies were that useless, how did we end up with DVD-Audio up to 192KHz? and why are DVD sampled at 48KHz instead of 44? Why did dts released a dts 96/24 standard?

Jojo, I had only 5 tracks on the CD, but my friend didn't even hesitate. Here are his comments while listening to the CD (and I was there with him):
TR1: Beuu, this is horrible
TR2: This is fairly good, but something is missing
TR3: This is the original
TR4: This is fairly good but not as good as the second one.

And effectively, the order of the tracks was: 160 256 orig 192.

So I don't think he guessed anything, but I might be wrong. Remember this took place in less than a minute with me at his side!!!!!

What version of lame did you choose?

You use 44kHz AudioCds as input, so that doesn't matter. Just take preset standard and I doubt he will recognize it.

You seem so sure of yourself. A bit like I was when I came into my friend's living room with my CD. Maybe some people have better ears than others. And then again maybe not, and it was my mistake.

I didn't talk about it, but I already did a test like this one a few years ago with BladeEnc, and similarly a friend of mine did detect the 256kbps easily.

Oh well... I guess nothing is perfect.

It's quite possible that he could hear the differences. That's not in dispute. But the settings you used to make the mp3's are obsolete. People are just asking you to try this again, this time with MP3's created using the recommended alt-preset settings.
You've basically turned up on the board here and said the -R3mix settings aren't transparent. Well, nobody here is arguing with you! Those settings are not recommended for use anymore, and for good reason. You need to remake these mp3's with a recommended version of LAME, on the recommended VBR alt-preset settings.
Then come back and tell everyone how you get on.

I'm not so sure, but as JJ180 already said, your settings are obsolete. So all I asked you is to try again with better settings which have undergone heavy testing including many blindtests contrary to -remix, which only was a line of settings.

and blade is known to be one of the worst mp3 encoders, even up to 320 kbps

Not only obsolete apparently, but also buggy.
The lame encoder used for this test ("the latest lame at the time (summer 2004)") had some troubles with -q 0 mode, which lead to terrible artifacts in some case and at lower bitrate. I can't say if those artefacts could be heard at higher setting, but the fact is that -q0 with previous lame was not a safe choice.

BTW, guruboolez, how do you think the new alpha is looking up?
I have been meaning to try it myself, but I am away on vacation.

Do you mean the seventh alpha? I didn't tested it yet. Probably at the end of this week. But if quality didn't regress compared to previous alpha, I would say that this release is the most enjoying version of lame since the release of 3.90 gold

Blade is more-or-less evil. Don't let it near your audio.


Marketing mostly. I think for 48khz there is a more reasonable rationale based on being evenly divisible into the sampling frequency 96 or 192khz masters (somebody with more audio processing knowledge may be able to fill in the reason, or tell me why I'm wrong).

I remember back when the Soundblaster AWE 32 came out - it was marketed by computer stores and the like as "the first 32-bit sound card!" Creative didn't make that claim of course, but they did little to dissaude people from the notion. At time having 32 midi channels was hardly the sort of thing you would be able to base a marketing strategy around.

IMO, if you use a current version of lame with --preset standard (or a higher preset), the differences in normal music should be inaudible, even under quite good conditions with an excellent pair of ears. There are a few sorts of sounds that will produce audible artifacts even at --preset insane, but they are quite rare. Even if there is an audible problem, it's not gonna be the lowpass.

Its the old misunderstanding:
bitrate is not unlimited in an mp3 - but you're asuming that it is when you think about the lowpass. When you increase the lowpass, then something else has to be encoded at lower quality.

About the DVD-A thingie - forget that. Thats marketing-talk. DVD-A is not about quality, but about multichannel and DRM. The increased bitdepth and samplerate is only there for the placebo-people.

But even asuming that one could hear a difference (i have yet to see anyone prove me that he can hear a difference with normal music - with both samples being from the same mastering), then with lossy codecs thats completely irrelevant, because the bits used on improving the ultra-high-freqs(something which is barely noticable) are taken away from the mid-high frequencies(something which is very noticable).

When you increase the lowpass, then you trade an extremely minor evil, for a much greater evil. In short, increasing the lowpass lowers quality.

Use the defaults. It's your job to tell the encoder which target-quality you want - and the encoder's job to decide on the best way to deliver that. NOT the other way around.

- Lyx

Well, while I agree with you on the concept, I disagree on the fact that moving the lowpass up automatically decrease quality. If this forum is right and 320kbps LP19.5 is well over transparent, then increasing the lowpass will decrease the number of bits allocated for the freqs below 19.5, but it was well over transparent, so it will still be over transparent. But you have a little more frequencies.

So all in all, if anything, it is a balance between transparent (as in enough bits) and transparent (enough freqs). Matter of choice at best...

Thanks

No. It is a matter of TRANSPARENT= cannot be told from the original, WHEN LISTENING TO IT.

Try the presets. Specially V2/aps and above. Almost no one can hear a difference. See if you are one.

To make you think just what you thougth there. 192khz must sound better then 44.1 khz cd.

192is effective if you are doing editorial work on the music as it will reduce artifacts introduced by rounding errors.

but as a final listening source. the difference is very small to non at-all (for 99% of the people)

Where do you get these statistics from?

Yes, that statisic certainly overestimates the number of people who would be able to tell the difference. Has there been a single person -- ever -- who was able to ABX a 192khz recording from a copy resampled (with proper techniques) to 44.1khz?

I think 96kHz/24bit was ABXed against CD quality 44.1kHz/16bit. The thread is here.

The thread talks about the setup for the test, but doesn't have any results as far as I can tell. Also, it's testing two variables at once (sampling freqency and bit-depth). It would be hard to say for sure if it's one or the other or both that someone is hearing (though I suppose audible quantization noise would sound like noise, and cut-off frequencies would sound like something else). Not that it's an invalid test or un-useful test, but it wouldn't determine if it was the lower sampling rate is what they're hearing.

IMO, the case for 24-bit is better than the case for 96/192Khz, but I don't have evidence on hand to back up that opinion.

Sorry the results are at this thread.

I'd probably agree with this. At the same time 96kHz file size would be more than double than that of 44.1kHz, yet the size for 24bit file would be 3/2 of the size of the 16 bit resolution file.

it will certainly not increase quality but is most likely to decrease it...what do you need the frequencies for anyway? You can't hear them...if you don't believe it, perform a blind ABX test (search the forum) and see if you can tell the difference between --alt preset standard (which lowpasses at 19khz) and the original CD. Have fun

About DVD-A.
Audio hardware has to lowpass sound to filter out high frequency quantisation noise. The process is usually done in two steps. First sound is resampled to the higher frequencies and a digital lowpass filter is applied. Then an alalog output is lowpassed again by an analog filter. This is done because there are no exellant analog filters. So the DAC of most good hardware is allready near 192 Khz. Nothing to change.
Introducing a 192Khz hardware increases quality because it decreases the convertations on digital sound processing.

I won't debate that, but has anybody ABX'ed 96kHz/24bit against 48kHz/24-bit?

Personally, I'd prefer at least 96kHz sampling rate in a digital format, to make the signal reproduction closer to that of analog. Yes, a sampling rate of at least 4× the frequency range.

I don't know about ABXing, but I can tell you for certain that at 44.1kHz sampling rate, you can record a 22kHz tone, but you will have no idea by looking at the digital data produced whether it's a sine wave or a sawtooth wave or what, just that it's a multiple of 44.1kHz. Likewise, how the DAC on a player will reproduce the signal from said data is equally unreliable.

Distortion can make high frequency sound fall into the audiable range. See KikeG's post here.

But there's no ambiguity at all - a 22kHz sine wave is a 22kHz sine wave, and that's what you'll get. A 22kHz <any other> wave has frequency components way above 22kHz which will be removed by the low pass filter, hence you'll get a 22kHz sine wave at the output.

Correctly implemented 44.1kHz sampling does exactly what it says on the tin: preserve everything below half the sampling rate. Nothing more, nothing less.

Cheers,
David.

I did a small test once to find my lowpass threshold.

I used the samples of Mustang Sally provided by ff123 here. I was able to ABX successfully only up to 14kHz

I know it was just one specific sample but from that day forth i stopped worrying myself with lossy encoder's lowpass. I mean LAME @ 128kbps lowpassess at around 15kHz...

And this is how it's meant to be done! I salute you.

If only more people were to follow...

Well, here's one more... :-)

Because of this post, I found these test samples and tried it myself.
I could not consistently perform a successful ABX beyond 14kHz either. (I nailed it a few times with the 15kHz sample, but that was probably dumb luck.)

This was probably said a couple of hundred times before, but...

1) Low pass filters -- you need good equipment (hi-fi sound card / hi-fi earphones) to hear anything above 16 kHz at all, and with a typical pair of desktop speakers, you won't hear anything above 12-14 kHz.

2) Distortion at high bitrate is possible, even with 320 kbps. Like was mentioned here, lame 3.94/3.96 had some bugs with -q0 (drove the masking thresholds way down). But even if masking is set properly, there are still samples that "fool" psychoacoustic models. Check out fatboy.wv from the original GPSYCHO quality page (http://lame.sourceforge.net/gpsycho/quality.html). That sample blows away just about every psymodel out there, GPSYCHO and NSPSYTUNE included (they think the noise is masked but it's still quite audible); the only way to encode that decently with LAME is --athonly at 320 kbps CBR.

3) Because of (2), (isolated) artifacts are normally much more common in VBR modes than in high bitrate CBR. In CBR, there is always a set minimum amount of bits to encode a frame, and the encoder will use it all even if all noise is masked (giving you a good margin of error in case the psymodel trips up). VBR modes rely on the psychoacoustic model to determine the minimum amount of bits to encode the frame. The idea is for all frames to have constant audible noise with minimum bits possible -- but in practice, there will still be passages which will use much more bits than they need to, and there will be passages which use too few bits because the psymodel thought all noise to be masked and it was not. The former is not the problem -- it only increases bitrate. The latter results in artifacts. That's why people (and presets) use minimum bitrate switches for VBR -- if for some reason the psymodel decides to use 56 kbps frame for a complex part, -b128 will ensure that at least 128 kbps frame is used. But of course this only masks the problem, and has a nasty side effect of driving up the average bitrate.

Or you can simply drive the masking thresholds all the way down, so even those parts where the psymodel was wrong will still sound decent. This is what -V0 does. Of course, then you have a whole slew of frames using way more than they need, so this also only masks the problem.

Or you can adjust the masking thresholds dynamically based on perceptual entropy as calcualted by the psymodel. LAME does this in all VBR modes, any presets. For the most part it works, but this does not take into account the frequency domain information.

In other words, there is a lot of jury-rigging in LAME VBR modes; they're probably the most tuned ones and so the optimal settings for a particular application are easy to pick (--preset everything-but-the-kitchen-sink), but VBR modes are by no means the most fool proof.

Edit: One other thing is, I haven't yet had a chance to dig into the psymodel code in LAME, but from what I understand both use some sort of tonal masking (higher harmonics shadowing lower ones, vice versa?), at least GPSYCHO (NSPSYTUNE may use intensity based masking instead, but I'm not sure -- that's what I heard and that's what it sounds like). Anyway, the main idea is, of course, to exploit redundancies in the sound, either in the frequency domain or in the time domain (or both). But when you have some audio that sounds suspiciously like white noise, as the one you tested with, with all the sharp attacks which trigger the use of short blocks which use more bits than regular ones, this kind of audio is actually the hardest to encode because there is not much redundancy to exploit! (See applaud.wav test case on the original GPSYCHO quality test page.) I know it sounds funny, but when you just have a couple of violins playing, you only have a few principal harmonics and the rest of the sfb's are just silent, so the encoder won't use any bits for them -- and you can probably ATH away most of that annoying tuba in the background. So all the presets you tried may have worked perfectly fine with some Mozart piece, but with Biohazard or Slipknot, you find yourself in trouble! This is indeed where VBR comes into play, but it's not perfect either.
[EDIT: Upon further researching the LAME source code, this last piece about encoding noise is wrong. The more wide-range the noise, the more sfb's are filled -- but at the same time, the more noise can be introduced to the encoding without the listener noticing it. So random noice is not the worst case scenario for LAME's psymodel, albeit monotonic sound (or one with relatively few harmonics -- piano, flute -- not violin) is easily the best case scenario for MP3 in general.]