24bit vs 16bit, the myth exploded!

[esldude]

   
As sound pressure, not music.

regarding hearing 25 khz.  They heard 25 khz, 24 khz sine waves etc etc at levels with a threshold of 102-107 db if my memory isn't faulty.  So I am assuming they could hear it as a part of music were music to have something up there and nearly nothing near it.  I would think it takes very little in lower frequency content to mask it altogether.  Of course the same is true of most middle age people in the 12-15 khz region. 

 

[RazorJack]

  The devil is in the details, which seems to me to be the crux of the matter

 
+1 - actually entertaining that such a group of people who pretend to know signal processing to the point of arguing over equations that have nothing to do with the reality of the topic don't even recognize where their misplaced theories fall apart in applied science.  it is good though to see at least one other true signal processing guy around

 
Show me one scientific test where you (or anyone) discerns CD quality of an album from >16 bit, >44.1 kHz sampling frequency.
 
Let me guess, not worth your time?
 
The devil is not in the details. The crux of the matter is between your ears.

 

[groovyd]

   
Show me one scientific test where you (or anyone) discerns CD quality of an album from >16 bit, >44.1 kHz sampling frequency.
 
Let me guess, not worth your time?
 
The devil is not in the details. The crux of the matter is between your ears.

 
lol

 

[magiccabbage]

what a thread

 

[ab initio]

Look, if you guys want to move the discussion forward in a constructive fashion, you can start by contributing scientific principals and linking to sources. The best sources are places like wikipedia, wolfram, and free online webpages from universities, etc. Claiming you have some textbook that somehow invalidates the sampling theorem is not helping. You have to provide the arguments to support your claim---in this case math, since you are arguing against a mathematical theory---and provide third-party sources that establish your arguments as valid. People are here to learn, so they are very receptive to intelligent arguments. Until you provide said intelligent argument, you've got no basis for claiming the sampling theorem is invalid. Instead, you are either a) providing a lot of evidence that you don't understand the concepts yourself, b) arguing a strawman, or c) trolling.
 
You are nay-saying a well established theorem, but refusing to actually explain anything in a coherent fashion, nor provide any sources that back your claim. If your argument has any merit then it would prove invaluable to support your argument with evidence. "I have a MS in computer engineering" is not evidence. If you would like learn more about how to refute an argument, I encourage you to read the article linked in my sig. Furthermore, there is another thread linking to an explanation of common fallacies, so now you know how to avoid making statements and arguments that fail to further this discussion.
 
If you want to make a point about the Nyquist-shannon sampling theorem being invalid, you could go a long way toward supporting your claim if you can point us to a proof, example, or explanation on what breaks down. You might (or might not) find a publicly available source here.
 
Cheers
 
PS for those who are interested, the MIT course notes on Discrete-Time Signal Processing is freely available online, and is given by Dr. Oppenheim.

 

[bigshot]

The bottom line for them is "There's no reason not to use high bitrates- hard drives are cheap, processors are fast." But it's also undeniably true that there is no reason to use them either. They can't point to anything that says there is an audible difference. That's the stalemate here. Some of us listen with our ears, and others care more about measurements on paper that are beyond our ability to hear.

 

[a_recording]

Thought you guys might be interested, I made a video about 24 bit audio drawing from the Xiph.org resources and others. It was this thread which made me interested in this topic in the first place!
 
There's lots of little inaccuracies in the video but hopefully will give people a good primer about 24 bit.
 

 

[Danz03]

All these arguments! Let's settle them with a blind test, download the 4 files from the links below.
 
https://www.dropbox.com/s/4wszs6k2s146c4g/GA%2024A.aif
https://www.dropbox.com/s/v78d6asrlckhwzf/GA%2024B.aif
https://www.dropbox.com/s/16xm59jjgwv6ds2/QR%2024A.aif
https://www.dropbox.com/s/5duud72ih4ftc6u/QR%2024B.aif
 
There are 4 sound clips from 2 songs, 2 different mixes from each clip, A&B. One of each is a true 24bit 96kHz mix while the other one is a 16bit 48kHz mix but up-sampled to 24bit 96kHz without dithering. Can you guys tell which is which? If most people here can get it right, then it proves that 24bit 96kHz is worth it, if not, then we'll know it's just a waste of space. 

 

[stv014]

It could have been 44.1 kHz instead of 48, as the latter is much less common for music. By the way, none of these seem to be real 96 kHz samples, as all of them have a brick wall filter at either ~22 or 24 kHz. Also, there are differences in the audio band that suggest that these might be different masters, rather than just the same "high resolution" sample in its original and 96/24->CD->96/24 converted version ?
 
For those interested in testing the effects of quantization and band-limiting, I have some older samples here and here. Obviously, the samples used can make quite a difference, and for the lowpass filtering a 96 kHz source file would have been better, but some will probably be surprised how much the sample rate and bit depth can be reduced until the degradation becomes clearly audible.

 

[Notus]

Actually its very difficult to test if there is difference between 24bit / 96khz and 16bit / 44.1khz in practice. In my opinion you would need at least 4 - 10 songs with 3x samples from each song. Recorded up to something around 40khz to make sure there is all of the ultasonic frequency's. Then you remove all of the frequency's above 20khz on one of the sample's. Second file would need to be altered so you remove all of the frequency's above 20khz but introduce random ultasonic noise to the sample. Last would be the unaltered sample. You would have to make sure every hardware in the chain is able to playback the highest ultasonic frequency's in the samples in order to do this test properly. Then you need to get a lot of test subjects with both trained ears and untrained.
I don't think you are able to test this by simply downloading a file from the internet since you cant be sure if the hardware used was able to capture everything needed for testing.
Also this would need to be tested with both speakers and headphones.
Just my opinion. But what do i know, i don't have the deep understanding on the subject that many of you have.

 

[stv014]

  Actually its very difficult to test if there is difference between 24bit / 96khz and 16bit / 44.1khz in practice. In my opinion you would need at least 4 - 10 songs

 
Feel free to upload samples you think would be well suited for testing. For practical (file size) and copyright reasons, they should be limited to 30 seconds length, but that is enough if you choose a section that is most likely to have audible differences (if any).
 

Originally Posted by Notus /img/forum/go_quote.gif
 
I don't think you are able to test this by simply downloading a file from the internet since you cant be sure if the hardware used was able to capture everything needed for testing.

 
Fortunately, the downloaded files can be analyzed to find out if there is any "useful" (that is, probably still inaudible, but not just noise or other artifacts) ultrasonic content without an early roll-off. Obviously, the transducers used for playback make a difference, as most tend to roll off steeply above 20-30 kHz at most, but then the people who use them should not worry as much about recordings lacking ultrasonics anyway.

 

[Danz03]

What do you mean they are not real 96kHz samples? What made you think I used a brick wall filter at 22 or 24 kHz? You can hear that high? And what do you mean there are differences in the audio band? I personally recorded these tracks and mastered them myself using 32bit floating point and made sure they have no post production afterwards. With one version mastered at 24bit 96kHz, and the other at 16bit 48kHz and directly up sampled to 24bit 96kHz. I used 48kHz instead of 44.1 because since I'm not using dithering, up sampling 48kHz to 96kHz would be much straight forward and would cause less error. And why would someone need 4-10 songs and recorded to 40kHz to test it properly? I thought most people who agreed that 24bit is better can hear a noticeable difference right away.
 
Quote:

  It could have been 44.1 kHz instead of 48, as the latter is much less common for music. By the way, none of these seem to be real 96 kHz samples, as all of them have a brick wall filter at either ~22 or 24 kHz. Also, there are differences in the audio band that suggest that these might be different masters, rather than just the same "high resolution" sample in its original and 96/24->CD->96/24 converted version ?
 
For those interested in testing the effects of quantization and band-limiting, I have some older samples here and here. Obviously, the samples used can make quite a difference, and for the lowpass filtering a 96 kHz source file would have been better, but some will probably be surprised how much the sample rate and bit depth can be reduced until the degradation becomes clearly audible.

 
 

  Actually its very difficult to test if there is difference between 24bit / 96khz and 16bit / 44.1khz in practice. In my opinion you would need at least 4 - 10 songs with 3x samples from each song. Recorded up to something around 40khz to make sure there is all of the ultasonic frequency's. Then you remove all of the frequency's above 20khz on one of the sample's. Second file would need to be altered so you remove all of the frequency's above 20khz but introduce random ultasonic noise to the sample. Last would be the unaltered sample. You would have to make sure every hardware in the chain is able to playback the highest ultasonic frequency's in the samples in order to do this test properly. Then you need to get a lot of test subjects with both trained ears and untrained.
I don't think you are able to test this by simply downloading a file from the internet since you cant be sure if the hardware used was able to capture everything needed for testing.
Also this would need to be tested with both speakers and headphones.
Just my opinion. But what do i know, i don't have the deep understanding on the subject that many of you have.

 

[ab initio]

One can use a spectrum analyzer abd other mathematical tools to check the validity of the tracks being tested. In this case, we want to test the difference between sample rates and bit depths, not differences between masters or software filters, etc.

One should start with the high resolution original and compare with downsampled versions of the same track. Special care must be taken so that the downsampling algorithm doesn't introduce aliased signals (e.g., one cannot niavely decimate)

Cheers

 

[stv014]

  What do you mean they are not real 96kHz samples? What made you think I used a brick wall filter at 22 or 24 kHz? You can hear that high?

 
Well, I am not the one claiming to be able to hear a difference, but these graphs show there are problems with the samples:
 
   
   
None of the files seem to have any content that looks like audio above 24 kHz, and one of each pair drops off already at the Red Book standard 22.05 kHz. To me, even the higher bandwidth versions look more like low level artifacts between 22 and 24 kHz from bad quality sample rate conversion than real content. The lower bandwidth files include what seems to be noise shaping in the ultrasonic range. I would guess the low bandwidth samples are resampled Red Book (which is indeed not uncommonly sold as "high resolution"), and the others have been converted from that to 48 kHz with not quite perfect resampling. However, the low frequency difference in one of the tracks is odd, at first I thought there is a mastering difference, but I am not sure.
 

  I personally recorded these tracks and mastered them myself using 32bit floating point and made sure they have no post production afterwards.

 
If you recorded them yourself, maybe the sample rate was converted due to some software problem. Unwanted sample rate conversion by the drivers or the operating system is a common issue on Windows. But why record twice, when you can just convert the higher quality version in software, which avoids unexpected differences like what is shown above ?
 

  I used 48kHz instead of 44.1 because since I'm not using dithering, up sampling 48kHz to 96kHz would be much straight forward and would cause less error.

 
For a good converter, non-integer ratios should not be a problem. These tests show 200 dB stopband (imaging) rejection achieved with resampling from 44.1 to either 88.2 or 96 kHz,  and the only practical difference is that the latter takes longer to process (but by a factor of less than 2).

 

[bigshot]

Whenever you put up listening tests on the internet, it's inevitable that people are going to cheat and use measurements instead of their ears.