Dithering analog summing and hardware inserts

So have I got this right for my work flow (as a best practice)

I record in Cubase Pro 9 as 24bit (44.1)

I mix out of Cubase using a Crane Song HEDD 192 as my A/D/A converter loop that feeds my hardware bus compressor than is then recorded back into Cubase at 24bit.

So the theory goes -

1. I should insert a dither plugin at Cubases main stereo out before it goes to the HEDD's D/A

2. I should set said dither plugin to be 24bit so I'm adding dither noise but not reducing the bit depth (which I don't want to do)

3. Do you turn off all the fancy noise shapping etc and just use the simplest dither possible?

4. No dither plugin needed on the A/D side of the HEDD returning to Cubaase.

5. Finally when I'm Mastering to create 16 bit files for say Spotify (via CD Baby) then I add dither again at the mastering stage and set it to 16bit for the final bit depth reduction. This part of dithering I can definately hear doing something!

I just want to make sure I have the principle correct.

Thanks

Safe to say it doesn't hurt to dither on hardware sends out of the DAW.  Do dither whenever reducing bit depth, or whenever the DAW is doing math on the audio. How many dB down is the dither, anyway?

Ok, I do have an answer for Logic Pro: it outputs 32-bit audio. I used Bitter, and a few different apps so I could send audio from Logic to another app, and host Bitter there. The apps were the menuBUS app, and Rogue Amoeba's Loopback plus Hosting AU. I tested both the master output and the I/O plugin, and the audio received at the next host was 32-bit, which is be the same audio that the interface would receive.

After that, it's up to the interface to deal with the 32-bit audio, which in RME's case is first done via TotalMix, their software mixer. Reference docs are here: www.rme-audio.de/en/support/techinfo/hdsp_totalmix_hardware.php --> "The output signal is truncated to 24-Bit without dither." They proudly defend the quality of their 24-bit fixed point architecture, it's a very interesting document: "This noisefloor - in reality quantization distortions - is generated by truncation at 24-Bit".

RME is a very opinionated brand (anyone that has checked their forums knows what I'm talking about) and they have decided that truncation distortion is not a problem when working at 24-bit.

Logic is doing the right thing here, letting the interface "decide" instead of truncating or dithering to 24-bit.

Now, I suppose a test could be devised to hear what happens with compounded truncation distortion. A short acoustic guitar recording, looped DA and AD say, 32 times, with and without dither, and then "mastered". Maybe someone has the willpower to do it, I don't...

That makes sense. So what's the dither actually doing?

Depends on what the processing did to the signal.  In layman's terms, during processing, a lot of "excess" bits are created that may or may not be usable audio data.  It's up to the coder to write the process to be smart enough to round or truncate the least useful parts of the data while attempting to retain the useful audio results.

Sometimes truncation won't result in audible issues at all.   Sometimes it will, depending on the coder's expertise and what exactly they are doing with the data.

It's entirely possible to retain the integrity of the original data while processing at a higher precision, but it's costly in terms of processing power and time, so it's another trade-off determined by those who code the process.

So those folks who are determined to say that that truncation/rounding always result in audible errors are incorrect.  Now, if they said truncation might result in errors, or if they said any processing inherently introduces errors, then they'd be correct.  

In any case, dithering always further reduces the precision of a signal, regardless of whether truncation did or not, but is done to hide the effects of the processes.   I like to think of it like a Picasso painting.  Up close, you see the mess of colors and stokes as overpowering mess.  You step back and the distance causes you to lose the focusing precision and now everything looks a lot more cohesive and all those colors and strokes blend into a masterpiece.  Same thing here, the (dither) noise "blends" the (processing result) noise and mess into something more aurally palatable. 

This also doesn't take into account whether or not the original signal was of sufficient quality to survive processing without significant artifacts.  You can certain have a 24bit encoded signal that only effectively uses 10bits of the A/D due to noise floor, etc.  This is called ENOB (effective number of bits) and is a more true representation of the quality of the encoded signal, but manufacturers have a love for SNR measurements as a quality metric due to their ease of conveyance but it tells you nothing about the quality of the actual encoded signal, only of the potential quality you can achieve. 

For low ENOB data words, pretty much anything you do will result in significant data loss as most of your bits are already noise.  High ENOB words can be processed more heavily without significant degradation. 

This is why some converters will sound better at 16/44 than others at 24/96 if you don't take into account such things like the aliasing/Nyquist filters, etc, and only use ideal models.  

However, those who say that "listening is the only way" are the most correct IMHO.  Since we are depending on the coders to be smart enough to write processes that waste the least amount of our initial data and we don't really have any insight into their work, we just have to listen to the results and choose what works best for us.

Arguing academics really doesn't do much.

Sounds about right. Different dithers do have different sounds, which ine you should use I don't know. I do know that Chris J (Airwindows) does have at least half a dozen different dither plugs available though, so there must be some reason that some people prefer oner over another.

The thing about Delta/Sigma DACs is that they inherently move quantization noise into the upper image bands when oversampling, but also need reconstruction filters on their outputs that both convert current to voltage and form a low pass element to reduce upper images.  These low pass filters effectively remove quantization noise from DAC outputs after it has been pushed into the higher bands.

Yet another reason that dithering is relatively un-needed for delta-sigma DAC outputs unless the data stream has been heavily processed in the DAW.

What dither does is randomize the bottom bit before the truncation so it can't chatter like a noise gate instead of just sounding like hiss. I wouldn't use noise shaping for 24 bit dither and no more than light for 16 bit because strong noise shaping can have a negative impact on lossy encoding and even potentially digital volume controls. It was a cool idea in the '80s and '90s when there were only CD players and they didn't have digital volume controls. Today's world is very different.

ADC and DAC chips are separate modules that have 24 bit fixed in and out in most cases. They use dither for their internal processing but JJ told be it would take as much as 10 dB. of additional noise to begin to mask truncation distortion.

I should mention how I learned about this stuff.

When I first realized I needed to know about computers back in the mid-'80s, I started attending weekly meetings of the Berkeley Macintosh Users Group on the UC campus along with monthly meetings of the Music Special Interest Group at Stanford University. In fact my future wife and I were the only two people who attended both meetings!

One evening in Berkeley I was talking about modems with the person sitting next to me and he turned me on to a back door into the internet via a local BBS. I went home, logged on using dial-up and set out exploring. This was before the web existed so you needed to enter a search term in the terminal and wait for a response. Jim Johnston of Bell Labs had set up a newsgroup called rec.audio. The discussion was super nerdy and way over my head but I learned a lot from Jim and a number of digital audio pioneers. After AOL opened the net up to the public, most of the real experts split but JJ hung in until Bell/AT&T Labs canned everybody doing audio and JJ was hired as the head of audio for Microsoft and needed to go under deep non-disclosure.

Bob Katz and I were part of that newsgroup although, sad to say, no DAW developers were there except for someone from Lexicon.

That's good to hear but I can't imagine why anybody would want to record less than 24 bits unless they were using a DAT machine, a CD recorder or an early video deck.

Just to be clear, I'm talking about the signal that feeds a DAC needing to be dithered.

I'm still confused about where the dither plugin should be for hardware external effects in Cubase.   With multiple external effect inserts on one channel, should all I care about is one 24 bit dither plugin in the first slot with no noise shaping, which only effects the DAC of all external effects in the channel insert slots?

Cubase uses a bus system for it's external effects. It also associates each external effect with a midi channel and there's a delay compensation mechanism. So, I don't know how that effects what dithering is trying to accomplish. There is a DAW "pre gain" that comes before the insert, but I rarely use it. I do some clip gain though. Would that qualify as "gain"? I have some confusion regarding "gain" and "volume" as it relates to the DAW and in the context you're using it.  Could you clarify? Are  fades considered volume or gain in the context you're using it?

I always have plugins before my hardware inserts. They qualify as signal processing in my mind, but I thought they remain 32 bit float in the DAW? They do have individual volume controls though, but again I'm not sure if that's neutralized by being in the 32 bit DAW environment.

I've never done this, but I'm going to do this one time to satisfy my curiosity. Thanks for your thoughts on the subject!