Exporting 24 bit masters from 32 bit Nuendo Projects

[David Lee]

I have a project that was recorded in 32bit float in Nuendo, Mastering house wants 24bit. On export, do I put the UV22HR set to 24bit on the master buss or simply export without it, set export to 24 and truncate the floating bits?

...admin, if this needs to be in the Music Computers forum, please move away..thx

[pingu]

Yes use dither when reducing word length.

[Lupo]

Floating can't be properly dithered. (link)

Truncate to 24 bits. The noise in the sources will drown the quantization anyway.

[pingu]

This may turn into a shitstorm.

[Kiwiburger]

It's a great question ...

Following Lupo's logic - no 32 bit floating point DAWs should ever apply dither.

I read the link - and it makes good sense. But I think there could be a misinterpretation here:

Obviously if you were creating a floating point file, it is not appropriate to add dither.

But we are talking about converting the floating point file into a fixed point file. I believe it is appropriate to add dither at that point.

Otherwise - why add it when creating a 16 bit file?

And why would dither plugins have a 24 bit noise option?

[David Lee]

Originally Posted by pingu

This may turn into a shitstorm.

i'm sorry, i really am...i just need to know!!!!

oh crap...i just pushed the red button that says 'do not push'

[David Lee]

ok. adding dither = adding noise...i understand that

is it necessary to add noise to a non real bit rate?

oops, i pushed it again...

[Lupo]

Had no hostile intents with the above post! It's just one of the quirks of floating.


Kiwiburger wrote:
>Following Lupo's logic - no 32 bit floating point DAWs should ever apply dither.

Someone may find a way to do it that removes enough quantization noise to be worth it. AFAIK, we're not there yet.


>But we are talking about converting the floating point file into a fixed point file. I believe it is appropriate to add dither at that point.

The problem is, where to add the dither? A steady noise, like regular dither, will not relate to the signal in the expected manner. The quantization noise of floating point varies with the level of the signal. Adding a variable level dither noise will not work either, as the dither then will be signal dependent, not random.

>Otherwise - why add it when creating a 16 bit file?

If you go from 24 bit fixed to 16 bit fixed, there'll be massive benefits of adding dither.

>And why would dither plugins have a 24 bit noise option?

It's technically correct to add dither whenever bit depth is reduced, as Pingu noted. Those plug ins who have the option to add 24 bit noise probably use a higher number of integer/fixed bits before reduction to 24 bit.


In this case, where there's only one quantizer process in the very last export of the project(32 float to 24 fixed), I thinkt the quantization noise will be effectively drowned by the noise floor in the sources in the mix itself.


David Lee wrote:
>i just pushed the red button that says 'do not push'

No way. It's a very good question!

>is it necessary to add noise to a non real bit rate?

Bingo!

The question can be restated as: 'is it possible to add a non-correlated noise to a non-real bit rate?'

IMO: no.


Best wishes,

Andreas Nordenstam

[bob katz]

Originally Posted by Lupo

Floating can't be properly dithered. (link)

Truncate to 24 bits. The noise in the sources will drown the quantization anyway.

Floating can be dithered if you know the level and you do it at that point on the conversion to 24 bits.

As to whetehr the noises in the sources or in D/A converters will drown out the quantization, I'm sitting on the fence on that. I think I've heard the most subtle sweetening of the sound when dithering from float to 24 bits and so I try to do it as a rule but I'm not uptight about not doing it. I am concerned about cumulative quantization distortion if you truncate each time in a number of stages, and dithering to 24 has NEVER hurt so far, much like chicken soup.

BK

[Cellotron]

fwiw -
my own opinion is that a single final truncation of a 32bit IEEE file - which in the majority of mixes I've seen just a 24bit file in a 32bit container anyway - to 24bit -
or dithering the same using the dither of your choice does not make a rats a$$ worth of a difference in terms of the final sound of your product.

Honestly - the least significant bit of a 24bit files is way below the noise floor. I simply don't think you are going to hear the difference between the 2 options - and I have a $100 bill I'm ready to put on the table in a bet to any "golden ear" to actually pick out which 24bit file is dithered and which 24bit file is truncated 7 times out of 10 in a blind a/b/x.

I think the real degradations happen only with repeated truncations or ditherings of 32bit files to 24bit as the returned sums are passed from one processor to another in a digital processing chain. The DAW I use, SAWStudio, mainly avoids these problems by having full 32bit integer sums returned by every process in the signal chain and then truncating (or having dither of your choice applied as an end user option) only once prior to outputting at 24bit.

It should also be noted that every single DA converter I am aware of needs a 24bit integer figure at output - so when you're playing back a 32bit IEEE file directly the software is either truncating or dithering it as you are monitoring it anyway.

Best regards,
Steve Berson

[David Lee]

so we're splitting hairs here....unless i'm recording in a silent chamber under very controlled conditions....which i'm not, it's a live show. a good sounding live show though.

thanks for the info guys.

[Lupo]

Hi!

Originally Posted by bob katz

Floating can be dithered if you know the level and you do it at that point on the conversion to 24 bits.

Doesn't that mean that the exponent in the floating signal have to be the same for every word in the track to be dithered?


Andreas N

[ed littman]

Originally Posted by Cellotron

Honestly - the least significant bit of a 24bit files is way below the noise floor. I simply don't think you are going to hear the difference between the 2 options - and I have a $100 bill I'm ready to put on the table in a bet to any "golden ear" to actually pick out which 24bit file is dithered and which 24bit file is truncated 7 times out of 10 in a blind a/b/x.

,
Steve Berson

I would agree.
do a test......take a 32bit file & render 2 files, one to 24 bit trucated & the other with 24 bit dither.

take both files & do a null test with the original 32 bit file.open your fft meter......you may be surprised at the result.
the truncated file will almost null perfectly . the other will show the dither noise. not sure if you can hear the difference but I just may go with truncation in the future.
Ed

[NeoVXR]

example: (I hope it is correct)
processing chain with 4 steps. everytime some float format result is truncated to 24 bit fixed.
without dithering, the errors would be correlated to the signal and add up. in a fictional environment there would be 4 times the error signal level.
with dithering, the errors are turned into decorrelated noise. there is a law about summing of independent noise sources. eventually we end up with twice the noise level in 4 processing steps, instead o4 4 times the error (distortion) level.
therefore dithering is recommended, the more processing takes place.

[bob katz]

Originally Posted by Lupo

Hi!



Doesn't that mean that the exponent in the floating signal have to be the same for every word in the track to be dithered?


Andreas N

I'm thinking yes, but I would like to hear more from a recognized DSP expert.

[bob katz]

Originally Posted by ed littman

I would agree.
do a test......take a 32bit file & render 2 files, one to 24 bit trucated & the other with 24 bit dither.

take both files & do a null test with the original 32 bit file.open your fft meter......you may be surprised at the result.
the truncated file will almost null perfectly . the other will show the dither noise. not sure if you can hear the difference but I just may go with truncation in the future.
Ed

In general, Ed is right. But you can MEASURE a distortion difference with a test tone. (between dithered or truncated). However, the distortion product is way down (the peaks of the distortion of a single tone on the order of -141 dBFS). And the products are rather ugly looking on the FFT, lots of inharmonics, if you look at some of the diagrams in my book.

But I agree that for a single truncation, this distortion is masked by all the noise sources in the system, including and especially the DAC itself. But as to whether cumulative 24 bit truncation is audible (as pointed out by Steve Berson and others in this thread), I'm on the side that it is slightly audible and should be avoided.

[nathanvacha]

Perhaps one of you bit-heads can answer this for me, or at least agree with or refute what I think is correct. I like the sound of taking a mix onto tape, even stereo out of the converters to the deck and dump it back in. If I record onto tape from say 24/96 and dump back in a new file @16/44.1, am I cheating the rules of SRC and bit reduction or is it gonna stick with me...

Will it function the same as truncation? Will the tape provide a natural "dither"? And really it will come down to an auditory comparison, but is it a better trade-off to use high sample rates for plugs and reduce it this way, or record @ 44.1? Just a question for the brains.

[Farview]

Originally Posted by nathanvacha

Perhaps one of you bit-heads can answer this for me, or at least agree with or refute what I think is correct. I like the sound of taking a mix onto tape, even stereo out of the converters to the deck and dump it back in. If I record onto tape from say 24/96 and dump back in a new file @16/44.1, am I cheating the rules of SRC and bit reduction or is it gonna stick with me...

Will it function the same as truncation? Will the tape provide a natural "dither"? And really it will come down to an auditory comparison, but is it a better trade-off to use high sample rates for plugs and reduce it this way, or record @ 44.1? Just a question for the brains.

In your situation you are not converting the sample rate at all. At least not in the way most people refer to sample rate conversion. When you play the tape back into the converters at 16 bit, you are creating a new 16 bit file. You aren't converting a 24 bit file into a 16 bit one, so no dither is needed.

[ed littman]

Originally Posted by nathanvacha

Perhaps one of you bit-heads can answer this for me, or at least agree with or refute what I think is correct. I like the sound of taking a mix onto tape, even stereo out of the converters to the deck and dump it back in. If I record onto tape from say 24/96 and dump back in a new file @16/44.1, am I cheating the rules of SRC and bit reduction or is it gonna stick with me...

Will it function the same as truncation? Will the tape provide a natural "dither"? And really it will come down to an auditory comparison, but is it a better trade-off to use high sample rates for plugs and reduce it this way, or record @ 44.1? Just a question for the brains.

think of this process as if you were processing your mix in the analog domain, even if it's really a strait transfer. no dither needed, but it's common practice to convert to analog & capture at 24/44.1 or 16/44.1
Ed

[Lupo]

Hi guys!

Can't leave this thread alone with all this confusion still alive! =)

While I heartily agree that dithering is much prefered to truncation, there's a big difference between dithering an integer signal and a floating point signal.

Integer signals have a fixed noise floor that does not vary between each individual word. To dither is to add a non-correlated random noise to the lowest bits of this signal. This will trade the signal-dependent distortion from truncation for a steady hiss of signal-independent uncorrelated noise.

The big problem comes when the signal is floating point. The noise floor of each individual word varies as the exponent changes value. To add a random noise with flat volume to this signal will not dither it, as the dither noise needs to be confined to the least significant bits. Since the level of the floating point changes all the time, so does the level of the least significant bit! The noise will way more often than not be somewhere entirely else than on the least significant bit. This will add a layer of noise to the signal, but it wil not dither it.

If the level of the dither noise where to change along with the level of the floating point signal it would no longer be a random uncorrelated noise. The noise would depend on the signal and the point of dithering is lost. Dithering a floating point signal is simply very hard to do, if possible at all.

If the original is 32 bit float and you need 24 bit fixed output, the best solution would be to export a 32 bit or higher word lenght integer signal(without dithering) and dither this to 24 bits. If the output have to be confined to 24 bits, dithering may be worse than simply truncating it, as the dithering does not work on floating point signals.


Back to the original question:
>I have a project that was recorded in 32bit float in Nuendo ......

Analog->digital converters don't output 32 bit float. The signals you recorded was most likely 24 bit integer. If you only intend to record and export again, storing the files at 24 bit integer would be a better option as you won't have to worry about dither or conversion from floating to integer.


Andreas N

[bonne]

If you used an interger format plug-in with dither function, like one of the Waves plugs, wouldn't that solve the problem of getting the dithering right? Or would the Nuendo 32-bit float format overrule the plug-in interger format?

Jørn

[Lupo]

Dunno! If the process dithers to 24 bits from an internal higher bit integer operation, it seems reasonable to assume it should output a floating signal with the exponent at the same level on all digital words.

But the very act of writing 'resonable to assume' in the context of plugins and processing makes me wring and cringe.. :)


Andreas

[brandy]

That is interesting for me as well.

Nuendo has 32bit float internal processing, even when you have set the project to 24bit. (then the audiofiles you record will be 24bit, but the processing will be 32bit float).

So - i usually export 32bit float files - after mastering changing to 16bit using dither. Yo.

But what about exporting a 24bit file out of a 24bit project? (when the internal processing is 32bit float)

I recently made 24bit files out of a 32bit float project, i just used UV22HR dither set to "24bit" - that made me feel safe... without knowing if this is the right procedere. I checked reverbtails etc with headphone, i did not hear any artefacts in the fade out...



brandy

[Bob Olhsson]

I think the most important issue is making certain that the level of the floating point file doesn't exceed "zero" resulting in a clipped fixed-point file.

[Greg B]

Originally Posted by brandy

That is interesting for me as well.

But what about exporting a 24bit file out of a 24bit project? (when the internal processing is 32bit float)

I recently made 24bit files out of a 32bit float project, i just used UV22HR dither


brandy

And what about making a 32Bit Float file out of a 32 Bit Float project before mastering it? Is dithering a bad idea then? Should one only use dither when changing Bit rate, because I thought it could also be helpful after lots of effects processing (reverbs, EQ, etc.) when mixing it down?

As much as I read about this, I seem to get a lot of conflicting information.

[HotRats]

Originally Posted by David Lee

I have a project that was recorded in 32bit float in Nuendo, Mastering house wants 24bit. On export, do I put the UV22HR set to 24bit on the master buss or simply export without it, set export to 24 and truncate the floating bits?

...admin, if this needs to be in the Music Computers forum, please move away..thx

AFAIK you always get a 24bit singal after your master fader so I think you just have to bounce your mix (to 24bit) and let nuendo take care of the 32bit f.p. to 24bit conversion.

[bob katz]

Originally Posted by HotRats

AFAIK you always get a 24bit singal after your master fader so I think you just have to bounce your mix (to 24bit) and let nuendo take care of the 32bit f.p. to 24bit conversion.

That's not true. The workstation runs in 32 bit float. All calculations, including the master fader are done in 32 bit. The conversion to 24 bit takes place at the "interface to the outside world" or "when a file is made". The outside world would be a connection to a DAC, for example, or to an AES/EBU interface. At those points it gets converted to fixed point and then truncated. So if you want to make a genuine 32 bit float file at the output of the master fader, then you can. Or a 24 bit file, but it will be truncated unless you add a dithering module (plugin) at the output of the master fader.

Only if Nuendo explicitly dithers the signal to 24 bit fixed would it be dithered. I doubt this is done unless they advertise it... it's almost never done "automatically" in any workstation I've ever encountered. You can test this with a bitscope or an FFT. If you see the dither when you play a blank section, it's dithering, otherwise, it will be totally mute (no signal at all).

That said, at this resolution level, dithering may help, and couldn't hurt. But we already talked about that :-).

[HotRats]

You mean that during mixing what I hear in my monitor is a truncated 32bit float signal?

Well...always belived that the bit reducition was a normal and necessary oparation made by the sequencer before sending the signal to the DAC.

I feel like I have to ask you "are you shure?" but I won't do it because I respect you and you knowledge and I'm sure that you're shure :)

----------------------------------------------

just for information, the source of my belifes is the Protools whitepaper about the tdm 48bit -mixer. withe paper I tought I've understood

in the white paper the write

"Following the initial multiplication operation, inputs are summed and right shifted again. Then, the final result of the mixer—the output sent to the DAC or plug-ins on the Master Fader—is the dithered 24-bit result shifted left by the inverse of same amount as the accumulated right shift, thereby returning the original signal level to the DACs."

[Farview]

Originally Posted by HotRats

just for information, the source of my belifes is the Protools whitepaper about the tdm 48bit -mixer. withe paper I tought I've understood

in the white paper the write

"Following the initial multiplication operation, inputs are summed and right shifted again. Then, the final result of the mixer—the output sent to the DAC or plug-ins on the Master Fader—is the dithered 24-bit result shifted left by the inverse of same amount as the accumulated right shift, thereby returning the original signal level to the DACs."

We are talking about Nuendo, not Protools. Unless I'm wrong (very possible) but protools is 48 bit fixed. The dithering question hinges on whether it's fixed or floating point, not the actual word size.

[brandy]

Originally Posted by bob katz

That's not true. The workstation runs in 32 bit float. All calculations, including the master fader are done in 32 bit. The conversion to 24 bit takes place at the "interface to the outside world" or "when a file is made". The outside world would be a connection to a DAC, for example, or to an AES/EBU interface. At those points it gets converted to fixed point and then truncated. So if you want to make a genuine 32 bit float file at the output of the master fader, then you can. Or a 24 bit file, but it will be truncated unless you add a dithering module (plugin) at the output of the master fader.

Only if Nuendo explicitly dithers the signal to 24 bit fixed would it be dithered. I doubt this is done unless they advertise it... it's almost never done "automatically" in any workstation I've ever encountered. You can test this with a bitscope or an FFT. If you see the dither when you play a blank section, it's dithering, otherwise, it will be totally mute (no signal at all).

That said, at this resolution level, dithering may help, and couldn't hurt. But we already talked about that :-).

That makes sence - now it's "official" :-)

Thanks for explaining that again (and again ;-), that what you explained meets exactly my scenario.

brandy