dithering question

[Andreas G]

Hello,

this question bothers me for a while...

if you have a 32bit floating point file and you want to convert it to 24 bit,
you have to dither it – nothing new so far.

But what if you insert an analog device in a host application like Logic or Nuendo,
which works in 32bit internally? Wouldn't that also mean you have to dither before DA?

Any thoughts?

[Mr.HOLMES]

As far as I know dithering is the last step what the mastering house will do for you.
The reason for this is any later change in volume etc will make your earlier dithering useless.

[baikonour]

Originally Posted by Andreas G

Hello,

this question bothers me for a while...

if you have a 32bit floating point file and you want to convert it to 24 bit,
you have to dither it – nothing new so far.

But what if you insert an analog device in a host application like Logic or Nuendo,
which works in 32bit internally? Wouldn't that also mean you have to dither before DA?

Any thoughts?

if you have an analog device as an insert or buss etc..the signal send to this device from the DAW will have the resolution set in your converters. let say your project in logic/cubase is 32 bit but your converter ad/da works at 24 bit then what goes to the analog device is 24 bit therefore you will have to dither it to 24 bit prior to hit the d/a. BUT if your project is in 24 bit or less, the internal calculation will still be 32 bit but what will go out from the daw will also be 24 bit or less so no dither is needed.

[Andreas G]

Originally Posted by baikonour

BUT if your project is in 24 bit or less, the internal calculation will still be 32 bit but what will go out from the daw will also be 24 bit or less so no dither is needed.

I think this is really the crux of the question. Will it really be 24bit? E. g. in Logic you have a freeze function, a kind of temporary bounce or snapshot of tracks with all inserted effects to save CPU-Power. This temporary file will be saved in 32bit floating point. So if useing freezes together with 24bit-files one would have to dither (to the 24bit DAC)? Also is a freeze-file really a different thing (bit deep wise) than an unfreezed track?

Thanks so far!

[baikonour]

Originally Posted by Andreas G

I think this is really the crux of the question. Will it really be 24bit? E. g. in Logic you have a freeze function, a kind of temporary bounce or snapshot of tracks with all inserted effects to save CPU-Power. This temporary file will be saved in 32bit floating point. So if useing freezes together with 24bit-files one would have to dither (to the 24bit DAC)? Also is a freeze-file really a different thing (bit deep wise) than an unfreezed track?

Thanks so far!

That's a tough question! it will be interesting to know from the guys at Steinberg or Emagic, is there an automatic dither or because the resolution is so high, it doesn't matter?

[DAS]

I don't have much time to get in depth here, but there are a few problems. The first is that 32-bit float really has only 24 bits of precision in the data (25 really, but that's a technicality of sorts). The other 8 bits are used to scale that data. It's not much higher "resolution" per se. Second, dithering floating point data is a very tricky matter due to the scaling of the data. It either requires a computationally massive amount of dither (and swamps very low level signals) or the dither ends up being correlated to the audio to some degree, which is not good (there are other problems, but for now these two will suffice). I'm not sure what really happens in each and every case when "dither" is turned on in floating point systems and plug-ins, but I have my doubts about its integrity in some cases.

When you exit your 32-bit DAW to the analog world the signal is indeed put out at 24-bit (or whatever your converter needs to see) and then returned again at 24-bit to the system. Here, even though the signal may be dithered you're limited to the dynamic range of the converter anyway (not to mention the outboard gear) so I wouldn't be too concerned about the above.

Hope this helps.

[Andreas G]

Thank you DAS for your time! Though I'm not sure if I'm more confused than before.

Originally Posted by DAS

dithering floating point data is a very tricky matter due to the scaling of the data. It either requires a computationally massive amount of dither (and swamps very low level signals) or the dither ends up being correlated to the audio to some degree, which is not good

You are talking here more in a theoretical sense right? In reality I don't have really an influence on the amount of dithering, it's just a fixed noiselevel generated by a plugin I'd insert before DA to 24bit. So the question is does dithering in this case help to avoid truncation or not? Or are you saying I don't have to worry about truncation because there is no audible truncation?

Originally Posted by DAS

When you exit your 32-bit DAW to the analog world the signal is indeed put out at 24-bit (or whatever your converter needs to see) and then returned again at 24-bit to the system. Here, even though the signal may be dithered you're limited to the dynamic range of the converter anyway (not to mention the outboard gear) so I wouldn't be too concerned about the above.

Sure, but again the really important question for me here is , if I don't have to worry about truncation?

[DAS]

Yes, I was speaking more conceptually. I realize that in most cases dither is something you just turn on or off. Some applications let you choose different types of dither or instantiate special dither plug-ins. A purpose of dither is to help mitigate the potentially nasty side effects of truncation, and as long as it is implemented properly it's probably not going to hurt you sonically. By the same token I doubt it would hurt you sonically to not use it in your system due in large part to the way floating points systems can scale data.

The main thing I wanted to clear up (before going off on a tangent) is that converting from 32 bit float to 24 bit linear isn't throwing away much precision in your audio data. If you had 32 bits of precision and were headed down to 24 then dither would potentially be important. With or without dithering a floating point system should be perfectly capable of scaling the data to the proper 24 bit linear levels. This scaling is important because to some degree it mitigates the need for dither. In a 32 bit floating system your quantization distortion lives an effective 24 bits (144 dB) down from the level of the signal, even if you turn down the signal in the mixer. So when you think of it in terms of truncation versus dither you're sort of ignoring an important part of how floating point systems work. I wish I had more time to delve into detail, but suffice to say in most situations it shouldn't make a lot of difference.

I don't really know the actual implementation of dither in Logic or Nuendo enough to positively answer whether or not you should flip the dither switch on. On a practical level my suspicion is that it will not make any audible difference for you. Personally, knowing what I know, I wouldn't use it. Save that processing power for something that really makes a difference. But I would do some audibility tests before making that decision final. Try it.

[Andreas G]

Thanks again DAS,

though you have no time to go more in detail with the 32bit floating point concept,
which I totally understand, I've learned something from you!

[Kees de Visser]

Originally Posted by DAS

So when you think of it in terms of truncation versus dither you're sort of ignoring an important part of how floating point systems work. I wish I had more time to delve into detail, but suffice to say in most situations it shouldn't make a lot of difference.

Dither is very effective in decorrelating requantization errors in fixed point systems and its use should be highly recommended.
Some time ago I've compared 24 bit truncated and 24 bit dithered versions to the 32 bit floating point original and both suffer from correlated errors (at LSB level). The dithered version obviously has a higher noise floor. Since dithering doesn't seem to bring any advantage in this case, I prefer to truncate from 32 to 24.
With levels around -138 dBFS it's probably all rather academic anyway :)

[Andreas G]

Originally Posted by Kees de Visser

Dither is very effective in decorrelating requantization errors in fixed point systems and its use should be highly recommended.
Some time ago I've compared 24 bit truncated and 24 bit dithered versions to the 32 bit floating point original and both suffer from correlated errors (at LSB level). The dithered version obviously has a higher noise floor. Since dithering doesn't seem to bring any advantage in this case, I prefer to truncate from 32 to 24.
With levels around -138 dBFS it's probably all rather academic anyway :)

Thanks for your response. So I also did a test and so far I couldn't hear any audible degradation with the undithered file, even in very quiet passages. It seems indeed that the quantization distortion is lower than the converters ground noise.