32bit floating softens sound compared to fixed 24/32/48 bit

Grok · Post by **Grok** » Wed Sep 29, 2004 1:04 am

On 2004-09-28 22:33, astroman wrote:
(...)
In that sense: make music - if it sounds right, then it is right

my 2 cents, Tom

Agree; but today, with systems like SFP, "making music" is not what it was 20 years ago. Today, making music is also recording and processing audio files, and some engineering tasks. Understanding what it takes from a practical point of view doesn't hurt. More: it's the condition for versatility. There is also the possibility of being a pure musician with no needed audio engineering skills by playing violin or acoustical piano, or whatever.

If using a specific kind of distortion is wanted in a musical piece (distorted kicks, bitcrushed synths, aliased synths), it doesn't mean that the final mix of the musical piece has to be distorted or aliased, unless it is the composer's specific desire for the considered musical piece.

Considering the great work Creamware has done on their most recent synths, it really breaks my heart to deteriorate these sounds in an nondesired way. If it is wanted, why not, knowing that it is adding a digital PCM character to the intended analogue character of these synths, more or less. It can be used in a creative way.

Sure, music customers and radio listeners aren't always audiophiles, and the listening systems have generally many weaknesses. But a well recorded/mixed/masterized production will sound better than a poor recorded/mixed/masterized production on any systems, will it be an audiophile one or not.

So, "if it sounds right, it is right", sure. To have some knowledge for being able to properly record and process and not deteriorate by inadvertancy that musical sound which "sounds right" doesn't hurt.

Post by **hubird** » Wed Sep 29, 2004 6:18 am

nothing wrong with gathering knowledge and constantly checking for quality.
But this attitude may also lead to the conclusion that truncation from 32 down to 24 doesn't hurt or can't be heard at the end.

All I'm interested in is whether you can hear it or not, humanly spoken.

It'd be a time consuming job to test the summing effect of many single tracks being truncated during the process.
The differences, if any, will be extremely small, that's what I understand from this thread.

For now I stick with 24 bit

samplaire · Post by **samplaire** » Wed Sep 29, 2004 7:28 am

24bit/24h

astroman · Post by **astroman** » Wed Sep 29, 2004 7:35 am

you cannot hear it - there are no 32-bit converters on the card

from the precision point of view we're talking complete nonsense here.
check the specs of the best analog recorders - they have a signal to noise ratio way below the specs of a regular Scope card.
you'll hardly find a vinyl player with more than 70db - it's physically impossible due to the surface noise of the black disk

a mix will sound great because it retains the phase accuracy of the recording, but not because the signal has a higher resolution.

btw we've yet completely left out the weakest link in the chain: the speaker...

cheers, Tom

Mr Arkadin · Post by **Mr Arkadin** » Wed Sep 29, 2004 7:39 am

btw we've yet completely left out the weakest link in the chain: the speaker...

Not if you've got PMCs

Mr A

musurgio · Post by **musurgio** » Wed Sep 29, 2004 7:39 am

In am on Grok's side...
I really hear differnece when recording to Cubase from my SFP mixer where I add effects like Vinco and or Graphic eq, Compressors, Gates, Expanders etc when recording a live band consisting Drums Bass El gtrs Ac gtrs vocals.
Then when I hear the recordings I can hear the difference.
I can hear this small distortion.
It is because I was monitoring via SFP mixer when I was creating the sound of the instruments and then Cubase playing back these sounds...
Thats why on previous post I have always pointed that Cubase takes away something from the sound.
I use 32bit floating recording in Cubase.
I know true tape I use that too but that is for effect...
What would make a killer app for Pulsar is to make a major VDAT update that can use VST plugins adn midi recording /vst instruments.
I am sure this can be done.
Regards,
Dimitrios
Regards,
Dimitrios

astroman · Post by **astroman** » Wed Sep 29, 2004 7:54 am

of course you hear the difference - the difference between the math processing of 2 different audio engines

it's exactly the reason why I think Grok is 'bashing the wrong guy' - lot's of parts of the system have a more significant influence than the bit resolution.
But I don't think it's a question of being wrong or right - it's just about viewpoints and hopefully helps that everyone interested in those questions can make his very own most effective decision.
It's not exactly fun to keep huge files around for nothing, even if hd space is pretty cheap today...

cheers, Tom

musurgio · Post by **musurgio** » Wed Sep 29, 2004 4:14 pm

So dear Astroman,
You confirm that Cubase as other audioappz have their own recording/playback algorithm that introduces their own sound ?
I have tried several and found that Sawstudio has the best sound of all.
It uses 24bit fixed and I believe this has to do much with good sound as that it uses assembly language (machine code) for its program.
I have emailed Creamware about updating thgeir VDAT to a full sequencing program but they said there are no such plans...
Regards,
DImitrios

Grok · Post by **Grok** » Wed Sep 29, 2004 5:35 pm

On 2004-09-29 08:35, astroman wrote:
you cannot hear it - there are no 32-bit converters on the card
from the precision point of view we're talking complete nonsense here.
check the specs of the best analog recorders - they have a signal to noise ratio way below the specs of a regular Scope card.
you'll hardly find a vinyl player with more than 70db - it's physically impossible due to the surface noise of the black disk

a mix will sound great because it retains the phase accuracy of the recording, but not because the signal has a higher resolution.

btw we've yet completely left out the weakest link in the chain: the speaker...

cheers, Tom

I would like to make the distinction between two specifical forms of unwanted distorsions with an audio signal.

Distortions coming from weaknesses in the analogue parts of an audio system are one thing; digital PCM distortions are a complete different thing.

With SFP, the purpose of using 32 bit files doesn't come from the 24 bit AD/DA converters (if you use only these converters to record external analogue signals without further processing, then 24 bit audio files are the perfect choice); it comes from the digital PCM definition of the SFP synths & FXes. Truncating creates a digital distortion which is auditively very different from the previous known analogues distortions. Aliasing too was unheard before PCM. And aliasing can be heard even with the cheapest distorded speakers, isn't it?

It would be hasardous to believe that because the consumer analogue power amp and speakers have their analogue kind of distortions, these analogue distortions will auditively mask the digital distortions introduced by a misuse of digital PCM. The digital PCM distortions will not fade into the analogue distortions, they have not the same auditive print.
When recording audio files, these PCM digital distortions can introduce some "fragilities" in the audio signals that can be grown bigger as with a magnifying glass when doing further processing (mastering...).

Cheers,
Grok

Post by **hubird** » Wed Sep 29, 2004 6:44 pm

Grok, provide us a A-B test, two mastered mixes at 16 bit, dithered

I'd love a sound improvement by one click on a button, really

medway · Post by **medway** » Thu Sep 30, 2004 3:52 am

Remembering the Golden rule n°2 (dither has to be applied only one time), you see that you can't apply dither to all your recorded 24 bit or 16 bit files to minimize the truncature harmonic distortion when recording the files from Creamware's synths and FXes outputs. Dither is a noise, and we don't want to multiply this noise, we want it to be unaudible.

That's not necesarily true. It's generally better to always dither when reducing bit depth. The distortion caused by truncation can be worse than the added extra noise, but its up to your ears to judge. The idea that is had to be done last only applies to the mastering process. In other words you want the final dither from 24 to 16 to be the last thing you do in the context of mastering. Now like you said if possible its obviously better to not have to dither by recording at 32bit.

astroman · Post by **astroman** » Thu Sep 30, 2004 1:26 pm

On 2004-09-29 17:14, musurgio wrote:
So dear Astroman,
You confirm that Cubase as other audioappz have their own recording/playback algorithm that introduces their own sound ?

I cannot exactly confirm it, because I don't have Cubase myself.
Nevertheless it's a matter of fact that different developers have different programming styles (and preferences) - there's a magnitude of ways to design the exactly same process.
And since the math functions in this domain are highly sensitive to precision there's a good chance of noticable differences.
This has been confirmed here many times by Cubase users.

I have tried several and found that Sawstudio has the best sound of all.
It uses 24bit fixed and I believe this has to do much with good sound as that it uses assembly language (machine code) for its program.

it's not in the language itself, but anyoneone who's able to deal with that type of coding must be a very skilled programmer with an extremely high ability to abstract physical processes.
In that context I assume you're right with your observation.

Assembly language is frequently used to optimize the core routines of an application - often in a way that you edit a C-compilers output by hand for higher efficiency or increased precision.

I have emailed Creamware about updating their VDAT to a full sequencing program but they said there are no such plans...

... and they don't have the resources, unfortunately.

cheers, Tom

symbiote · Post by **symbiote** » Thu Sep 30, 2004 4:00 pm

Grok, the distortion you are referring to is called Quantization Error. It will happen on 24-bit recordings also, since sometimes the level of the signal you are trying to record is going to end up being between 2 quantization level.

For a 24-bit signal, the quantization error level is going to be around -144dB (so a 144dB SNR assuming your signal uses full range.) The best D/A converters that I know of operate at around 128/130dB SNR if I'm not mistaken, so the quantization error is going to be around 15dB LESS than the converter's noise.

Going from 32bit to 24bit is going to be pretty similar. You might gain a dB or two in SNR if you round instead of truncate, but the results won't be anything you can hear or perceive.

Applying gain during the mixing (thru effects, compression, or just level monging) and during mastering might amplify that error a bit, but you'll need to amplify it quite a bit (that's alot of bits) before the error gets randered on the best hardware around. Like I said, you can amplify it a whole 15dB worth before it equates the noise of the best converters around, so if you follow good recording practice, like recording at highest level possible, this is never, ever going to be a problem with this centuries converters.

Point being, there are going to be much higher noise sources in your signal path than quantization error. It's probably going to be on a par with your own ear's quantization error

Maybe if you have a nice Amek console with Neve converters and some freak-o sound system to go with it, you might end up perceiving something when playing stuff real, real real loud, but I for one would be too busy drooling over the console to notice a -144dB quantization noise error

astroman · Post by **astroman** » Thu Sep 30, 2004 5:14 pm

since truncation has been mentioned a couple of times: it's very unlikely that the last 8 or 16 bits are simply chopped off.
Probably the value is scaled first in a clever way to retain as much dymanic information as possible in the leading bits of the data word(s)

cheers, tom

symbiote · Post by **symbiote** » Thu Sep 30, 2004 9:31 pm

Pretty much. But it's basically still a decision between a 0 or a 1. Like, you use the 8 bits you chop off to choose the closest value to get the best representation. If you go completely random, you'll hit the right value 50% of the time. With a good algo, you can push this much higher, I guess close to 100% (more like in the 90%'s realistically without 1000 years of research and/or alien technology contacts (well, not really true, the algo is pretty simple.)) So you'll get the near-best 24bit approximation for your signal that exist.

Sometimes, you'll get 15432764 instead of 15432765 for some sample value, and you'll get a bit of distortion.

<font size=-1>[ This Message was edited by: symbiote on 2004-09-30 22:32 ]</font>

Post by **hubird** » Fri Oct 01, 2004 8:37 pm

On 2004-09-30 22:31, symbiote wrote:
Sometimes, you'll get 15432764 instead of 15432765 for some sample value, and you'll get a bit of distortion.

A great first sentence for a book