Test: 24 bit vs final CD product

[Jenz]

I've been trying to truly evaluate a final CD product vs the 24 bit mastered audio. Here's what I did:

I mastered all the tracks in Pro Tools (as I always do), recording to a new audio track that will contain one long file of the whole album, dithered and ready but still in 24 bit. I exported the file to 16 bit stereo interleaved, and continued with sequencing/master creation in Sonic PMCD. The slightly elusive PMCD has given me some headache over the years, so I figured I'd really test the outcome against my 24 bit file from Pro Tools. I had PMCD burn a CD, and then I used Pro Tools to import the CD back for comparison. So, basically I'm comparing the source against imported 16 bit versions that's been burned to CD (converted back to 24 bit, but what's truncated is truncated I guess).

EDIT: The source and obviously the destination SR is 44.1 khz.

I'm gonna post audio results of a 180 degree cancellation between the source and the CD outcome (on quiet parts as well as loud), but first I'm really interested in hearing what you guys think the difference could, should or probably will be? The music is rock/metal, slow stuff with many soft parts as well as really heavy.

[fenderbender.]

Originally Posted by Jenz

I've been trying to truly evaluate a final CD product vs the 24 bit mastered audio. Here's what I did:

I mastered all the tracks in Pro Tools (as I always do), recording to a new audio track that will contain one long file of the whole album, dithered and ready but still in 24 bit. I exported the file to 16 bit stereo interleaved, and continued with sequencing/master creation in Sonic PMCD. The slightly elusive PMCD has given me some headache over the years, so I figured I'd really test the outcome against my 24 bit file from Pro Tools. I had PMCD burn a CD, and then I used Pro Tools to import the CD back for comparison. So, basically I'm comparing the source against imported 16 bit versions that's been burned to CD (converted back to 24 bit, but what's truncated is truncated I guess).

EDIT: The source and obviously the destination SR is 44.1 khz.

I'm gonna post audio results of a 180 degree cancellation between the source and the CD outcome (on quiet parts as well as loud), but first I'm really interested in hearing what you guys think the difference could, should or probably will be? The music is rock/metal, slow stuff with many soft parts as well as really heavy.

Dither, shaping, and SRC will **** with the high-freq. content. It'll prb. sound more 'narrow' (as you'll loose a little HF). If you have a system that's of high enough fidelity to reproduce the differences between 88.2/96//24 and standard-res., you'll know immediately. If not, it's nothing to worry about.

[Jenz]

It fact, it completely nulls out. 100%, everywhere. NO difference at all. I was quite surprised to be honest, I mean people discuss the difference between dithers. At least this is a proof that my physical masters and PMCD are as good, or better than I hoped for (this was the reason for the test to begin with).

[Petrus]

Does not suprice me. I just calibrated my monitoring system to 83 dB SPL at -20 dBFS and also measured the noise floor of my mastering room: 30 dB SPL C. That means that the lowest detail in on a "perfect" CD would be 20 dB below the ambient noise which I would not be able to hear. On a 24 bit master (no such thing exists, 21 bits at best) the bottom would be 30 dB below the thresold of hearing. No wonder the difference can not be heard.

Dithering is around -70 dBFS at worst, and as the human hearing is not sensetive to signals around 18 kHz, those artifacts can not be heard in real life. Only if you turn up the signal real loud between the tracks, which is not normal.

What comes to loss of frequences above 20 kHz or so, hardly any material has any content there. Three conditions would have to be fulfilled before we need to worry about the loss: 1) microphone actually records above 20 kHz. 2) loudspeakers accuratelly reproduce signals above 20 kHz. 3) Listener can hear signals above 20 kHz.

I would bet that all those conditions are not fulfilled hardly ever. Only when we have a wealthy, young person who likes acoustical violin music.

[Audiop]

I don't think one can count on "perfect" masking from room ambient noise which is mostly in the low frequencies and also comes from all around us.

Also 83dB SPL at -20dBFS means peak levels of 103dB which is far from enough to reproduce many instruments. 110-120dB SPL is necessary for realistic reproduction of a wide variety of sources/sounds.

If you have a capable system the 16bit dither noise may be audible when cranking the volume with dynamic material.

But yes, I do agree that for most listeners this is not the case and the 16bit noise will indeed be inaudible.

What people need IMO is not higher resolution than 16/44.1 but better speakers and rooms and more powerful amps.


/Peter

[Petrus]

Originally Posted by Audiop

I don't think one can count on "perfect" masking from room ambient noise which is mostly in the low frequencies and also comes from all around us.

Also 83dB SPL at -20dBFS means peak levels of 103dB which is far from enough to reproduce many instruments. 110-120dB SPL is necessary for realistic reproduction of a wide variety of sources/sounds.

If you have a capable system the 16bit dither noise may be audible when cranking the volume with dynamic material.

But yes, I do agree that for most listeners this is not the case and the 16bit noise will indeed be inaudible.

What people need IMO is not higher resolution than 16/44.1 but better speakers and rooms and more powerful amps.

/Peter

Agreed, especially the last point!

But the fact remains that even 103 dB SLP peaks is louder than most people like (or can due to family/neighbour reasons) to listen. I might sometimes play at about 105-110 peaks absolute maximum, but only classical and not even that. Even then the dither noise is not there. Usually there is first the ambient noise of the venue, then microphone selfnoise.

[Nordenstam]

Originally Posted by Jenz

.. recording to a new audio track that will contain one long file of the whole album, dithered and ready but still in 24 bit. ...

I figured I'd really test the outcome against my 24 bit file from Pro Tools.

Herein lies the crux: the 24 bit file have the same dither as the 16 bit file. They should null! Try comparing a 24 bit file sans dither with a dithered 16 bit file.

Originally Posted by Petrus

But the fact remains that even 103 dB SLP peaks is louder than most people like (or can due to family/neighbour reasons) to listen. I might sometimes play at about 105-110 peaks absolute maximum, but only classical and not even that. Even then the dither noise is not there. Usually there is first the ambient noise of the venue, then microphone selfnoise.

As Peter noted, noise is not a total masker. Fletcher-munson curves tell us we're practically deaf in the low end at low levels. Real life dictates that most room noise is approximalely 1/f in distribution(as pink noise). So, most room noise is in a range where we don't hear it, making such numbers as used above a poor indicator of practical audibility.

This is the noise floor in my room in the most sensitive region, compared to pink noise:


That noise floor calculates to 21dB SPL A weighted(frequency weighting, not SPL meter speed). The spurious free dynamic range in the critical mid high area is over 120dB when adjusted for 83dB SPL = 0VU = -20dBFS (103dB peak). The same 120dB active range only reads 82dB when described in the typical dynamic range term from max peak to A weighted noise floor, even less so without frequency weighting. The difference between spurious free dynamic range and the normal "pessimistic" way of measuring dynamic range is typically somewhere between 3:2 and 2:1. That's why mediums with seemingly total lack of dynamic range still manage to convey a lot of music information. A noise floor that measures bad in time domain numbers can often be wonderful in the frequency domain - which is a lot closer to what we hear as humans. As wittnessed in a lot of good old equipment that can measure sorta badly while still providing a stunning real life performance.

103dB peak is not loud either. It's the bog standard NORMAL listening level. By cranking the RMS to 95dB, things do get loud, but then again, it's not louder than an acoustic performance with real instruments. If the recording is as dynamic as a real life event, peaks of such program reach 115 to 120dB. Which, in my room, gives a spurious free dynamic range at about 130dB.

With good noise shaping in a 16 bit file, spurious free dynamic range do approach 130dB in the most critical region. If my room had been quieter, or I had been using noise blocking in ears(as I did in some blind tests a while back), 16 bit would not be enough no matter what dither is used.

Summa summarum: The hearing system is a wonderful instrument that is capable of at least 120dB dynamic range. I see no reason to believe the ears suddenly loose that ability just because we're listening to music!


Best regards,

Andreas Nordenstam

[Jenz]

Originally Posted by Lupo

Herein lies the crux: the 24 bit file have the same dither as the 16 bit file. They should null! Try comparing a 24 bit file sans dither with a dithered 16 bit file.

Yes, the dither should null. But it also prooved that (in this case) there was no use to have the final product in 24 bit. I asked a few engineering friends of mine what they thought would be the audible difference between this 24 bit mastered version and the final CD version (without hearing it), and they used words as "slightly grainier treble", "a little less detailed" etc. As I believe many people around here would assume as well, perhaps without ever trying it. My personal conclusion is that most popular music would never gain from having a final product in 24 bit. And when people say that the CD format sounds shit, well, get a new CD-player.

[sedohr]

Not very scientific but when I have bounced a track to 16 bits and listen to the bounce, be it from a cd or straight out of the computer, I feel it sounds much shallower, lacks the juiciness it had in 24 bits. I hate it.
I can't put my finger on it, it's just a feeling. Which is what music is supposed to be about.

I've never done a null test on it but I'm very surprised that yours nulls out.
Well, every day you learn something new!

Kalli

[Nordenstam]

Originally Posted by Jenz

Yes, the dither should null. But it also prooved that (in this case) there was no use to have the final product in 24 bit.

24 bit with properly working 16 bit dither should be the same as 16 bit with the same dither. So you've basically tested two of the same files. As shown in the null test.

If you compare 24 bit without dither to the dithered version, 16 or 24 bit, you'll get the true comparison of 24 and 16 bit.

[Nordenstam]

PS: all 16 bit dither plugs I've tried also does the truncate to 16 bit after dithering. Have you checked the 24 bit file for activity below the 16'th bit? Chances are high that the 24 bit file container with dithering plug applied only contains 16 active bits.

[Jenz]

Originally Posted by Lupo

PS: all 16 bit dither plugs I've tried also does the truncate to 16 bit after dithering. Have you checked the 24 bit file for activity below the 16'th bit? Chances are high that the 24 bit file container with dithering plug applied only contains 16 active bits.

That's really interesting! I'll do a null test against a non-dithered version of the 24 bit file and report back. Could definitely be the explanation.

[w_stylz]

Also --- if you are not going to do any further processing -- 24 bit and 16 bit are not hugely different -- but if you are going to do further mastering eqing, you are introducing more error once you start boosting bands by 4-6 db .. If I remember correctly from science class, signifigant figures would be the terminology.

That all said, I would not be surprised to find a very small measurable difference with a null test but I also thing it would be spread out among the frequency response that it would not truly manifest itself until you started doing additional processing.

Otherwise, I would think that a 16 bit 44 snapshot would be pretty darn close since the additional detail stored in the file is probably well below the noise floor. You would have to probably have the music pretty cranked in an ideal listening environment and it would have to be very dynamic music I would think to hear a significant difference.

[Audiop]

Originally Posted by Jenz

My personal conclusion is that most popular music would never gain from having a final product in 24 bit. And when people say that the CD format sounds shit, well, get a new CD-player.

thumbsup


/Peter

[Virtalahde]

Shh.. 16bit/44k1 sounds inferior, the Internet says so!

[Petrus]

I have made a couple of test files from 24/96 masters which have random 16/44.1 sections* in them, and nobody so far has been able to tell where those "low-rez" parts are. Even if I listen with the timecode list in my hand I can not hear it.

Easy to see with frequency analyzer, though.

*) from the same 24/96 file to 16/44.1 with noise shaped dither, then backconverted to 24/96 and spliced in to the original 24/96.

[PatrikT]

Originally Posted by Jenz

I've been trying to truly evaluate a final CD product vs the 24 bit mastered audio.

Stop trying.

If the audio is designated for CD, the 16 bit version is the master and the 24 bit version is something for the archives.

Mastering is very much about checking if things are OK on/with the target media/resolution and very little about 180 degrees of cancellations against archived material.

Try tests like these with spinning turntables or a rotating tape machines to realize something.


Best Regards
Patrik

[Jenz]

Originally Posted by PatrikT

Stop trying.

If the audio is designated for CD, the 16 bit version is the master and the 24 bit version is something for the archives.

Mastering is very much about checking if things are OK on/with the target media/resolution and very little about 180 degrees of cancellations against archived material.

Try tests like these with spinning turntables or a rotating tape machines to realize something.


Best Regards
Patrik

Of course, I agree, but how often do you listen to the final 16 bit product while mastering? All I wanted to know was exactly what the difference between the two would be. And know I'm really calm about it.

[PatrikT]

Originally Posted by Jenz

Of course, I agree, but how often do you listen to the final 16 bit product while mastering?

All the time as long as the goal is CD-resolution.


Best Regards
Patrik

[DynaForte]

A couple points.

The fact that the polarity reversal cancels everything out doesn't show much.

It seems like you're saying it "proves" that the two signals are identical.

The question ith 16 versus 24 is if anyone can tell a difference in an ABX test.

Of course the signals will cancel - polarity is a reversal of phase and the two signals will still have the same level at all times.

Besides - WHERE is the polarity switch taking place?

You could not have taken BOTH of these files and put them into PT without converting one, which means they were BOTH 24 or 16 bit!



I do not believe that humans can actually pick out an ABX test. I don't believe people hear the difference in 24bit versus 16bit.

But, you didn't actually test anything.

[puremastering]

I Would be curious to know what Bob Katz think about it...

[bob katz]

Originally Posted by puremastering

I Would be curious to know what Bob Katz think about it...

I can't give any response on this question in a forum. How about you first read 3 chapters in a book on this subject I could recommend you, and a very good think session after reading that first? It's not fair to ask me to summarize pages of thoughts in a forum answer.

BK

[fourvector]

Originally Posted by DynaForte

A couple points.

The fact that the polarity reversal cancels everything out doesn't show much.

It seems like you're saying it "proves" that the two signals are identical.

The question ith 16 versus 24 is if anyone can tell a difference in an ABX test.

Of course the signals will cancel - polarity is a reversal of phase and the two signals will still have the same level at all times.

Besides - WHERE is the polarity switch taking place?

You could not have taken BOTH of these files and put them into PT without converting one, which means they were BOTH 24 or 16 bit!



I do not believe that humans can actually pick out an ABX test. I don't believe people hear the difference in 24bit versus 16bit.

But, you didn't actually test anything.

I don't get what your point is about the files both having to be the same bitrate by virtue of being in the same environment...

Pro Tools runs at 48 bits. If you give it 16-bit audio, it'll add 32 extra zeroes to each sample; if you give it 24-bit audio it adds 24 zeroes. They're still intrinsically 16 and 24-bit audio; the difference is that the 24-bit file gives 8 bits of meaningful data where the 16-bit just has zeroes; this difference is what the OP expected to hear when subtracting one from the other.

However on further consideration it does make some intuitive sense that the difference would be inaudible, since the dither is applied at the 16th bit, putting all the data in bits 17 to 24 beneath the noise floor.

I decided to try it with truncated 16-bit audio (no dithering) to see how it sounded. Surprisingly it also came out with mostly just a tiny bit of white noise...I guess since at higher levels the quantisation error amounts to a random deviation on each sample. I did hear some crumbly artifacts when the signal dropped below roughly -60dB or so.

So the way it seems to me is; the signals aren't identical, but the differences that do exist are predominantly noise, so that's why you don't really notice anything.

[Nordenstam]

Originally Posted by DynaForte

I do not believe that humans can actually pick out an ABX test. I don't believe people hear the difference in 24bit versus 16bit.

That's true for most end listener scenarios. Not all.

16 vs 24 bit with music originally recorded to tape (!): 24 vs 16 bit not audible? (10/10 blind test here: 24 vs 16 bit not audible?)

Another 16 bit vs 24 bit: Science Versus Gearslutz

Those links are not the end all of such testing, far from it. But they do give indications that at least some individuals can under some circumstances hear a clear enough difference to blind test it. The perception of TPDF digital 16 bit noise floor was possible even on a tape source recorded to digital at 24 bit. Presumably with an inherent noise floor measuring far worse than -90dB. Good shaped dither at 16 bit seems to be too difficult to distinguish from 24 bit on any practical music example. With lone sound sources at normal not-loudness-war levels, any 16 bit dither can be audible.


Regards,

Andreas Nordenstam


PS: would love to hear JJ's take on this subject!

[bob katz]

Originally Posted by fourvector

So the way it seems to me is; the signals aren't identical, but the differences that do exist are predominantly noise, so that's why you don't really notice anything.

The difference between a 24 bit source and 16 bit dithered result are EXACTLY the dither noise, no more, no less.

The Dither noise and the "music" signal add together EXACTLY as they would add in the analog domain. When I get a chance I'll illustrate this in Spectrafoo and put it up for you all.

BK

[lucey]

Originally Posted by Jenz

but first I'm really interested in hearing what you guys think the difference could, should or probably will be? The music is rock/metal, slow stuff with many soft parts as well as really heavy.

I find the differences between 24 and 16 (which are at -92 or so) are audible in the dense material as much or more than a single, pure track. For example, if something is darker with a lot of low end and compression, there can be a threshold where the 16 loses some of the depth/life and gets phat on the bottom. The eq adjustment to equalize might be -0.2 at 60, Q 2.5.

It's really about each case, no rules here. Sometimes the difference is not much of an issue.

[bob katz]

Originally Posted by lucey

I find the differences between 24 and 16 (which are at -92 or so) are audible in the dense material as much or more than a single, pure track. For example, if something is darker with a lot of low end and compression, there can be a threshold where the 16 loses some of the depth/life and gets phat on the bottom. The eq adjustment to equalize might be -0.2 at 60, Q 2.5.

It's really about each case, no rules here. Sometimes the difference is not much of an issue.

And this difference will also vary depending on the type of dither that you choose, which will mask (or unmask) different frequencies and low level information in the 24 bit source. You can't fit 24 pounds in a 16 pound bag, but you can get pretty darn close!

[Key]

Originally Posted by bob katz

The difference between a 24 bit source and 16 bit dithered result are EXACTLY the dither noise, no more, no less.

The Dither noise and the "music" signal add together EXACTLY as they would add in the analog domain. When I get a chance I'll illustrate this in Spectrafoo and put it up for you all.

BK

Hey Bob,
The thing is I am not seeing this as any sort of distinction from truncation. Just that with 16-bit truncation the noise/distortions are a little bit higher with an ideal test tone and a little more distorted. It seems to me it's not so much about preserving the original signal as it is picking your poison in regards to what type of noise you want to incur. And once you get to 24-bit I am thinking that the tradeoff is clearly in favor of truncation over dithering. But my experience might only hold water with a certain DSP depth - I am really not sure.

[bob katz]

Originally Posted by Key

Hey Bob,
The thing is I am not seeing this as any sort of distinction from truncation.

Subtracting the 24 bit source from the 16 bit dithered result is not the same as evaluating the results of plain (undithered) truncation!

Here's the kind of test you'd need to visually see the results of truncation (I haven't had the time yet to make one for you):

Take a low level test tone, say a dithered 24 bit sine wave at -60 dBFS. Look at this on a high resolution FFT. You should see no distortion products and a 24 bit noise floor. Now, truncate this to 16 bits and remeasure. You should see distortion products.
Dither it (the 24 bit dithered sine wave) to 16 bits and you should see a new 16 bit noise floor but no distortion products.

To repeat, the subtraction (null) test is NOT the same as testing truncation!!!!!! If, however, you take a truncated, distorted (not dithered) 16 bit file and subtract the 24 bit file from that, what will be left will be the distortion products, and you can see that on the FFT and you can hear the distortion residual on a pair of headphones though in some rooms you may have to dial up the monitor gain slightly. The kind of distortion residual that can result from different operations (e.g. compression, equalization) can vary considerably, and the distortion residual from the simple sine wave test tone will be rather "benign".

BK

[Key]

See I am throwing every test I can at this. When I use real world recordings it seems to do much better - self dithering? When I throw a sine wave test at it is when the truncation does poorly vs the dithering. And really it's still way down at the bottom. And once I hit it with a standard 60Hz + 1kHz IMD tone the results move closer together.

I tried a null test with 2 of my demos and on a fade out of a song. I tried using RMAA, HolmImpulse (but I don't know if I trust these results yet), and I am not seeing anything really effecting the main content of the file. It's just that the distortions are higher when I don't dither to 16-bit. But like I said the added noise in dithering to anything higher than that seems to be higher than the distortions incurred.