Stereo Dithered Mixer

[LUpton]

Isn't PT9 (non-TDM) using the 32-bit floating-point DAE engine from LE/MP?

Doesn't that render discussions of mixer internal dithering moot? I don't mean dithering for output bit rate reduction, just internal workings of the mixer.

...or am I full of it?

Larry

[daeron80]

Larry, you're correct. "Regular" PT9 (and isn't there a PT9 MP now? ) uses a 32-bit floating-point engine, HD Native uses a 64-bit float engine, and TDM-enabled HD uses a 48-bit fixed-point engine. Since floating-point data can't be dithered by definition, this thread only applies to non-Native HD, which is what the OP appears to be using.

[Greg Malcangi]

Originally Posted by daeron80

That's all true. But dither is noise. Granted, it's special noise in the sense that it's shaped to avoid being heard. But any noise in the vicinity of white causes dithering to take place. If you got (whitish) noise, you got dither. So if the noise in your tracks and mix is high enough, it will decorrelate truncation errors. If dither is below the noise floor of your mix, it will be buried by it and make no difference. So says Nika Aldrich, anyway, in Digital Audio Explained, ch 9, if I understand him right. And it fits what I've experienced over the years.

Kind of ironical that you should quote Nika, as I had quite a bit of discussion with him about a very closely related issue about a decade ago! I really don't want to get into all the detail again and the discussion was regarding ProTools of many years ago and is probably not applicable anymore. Here's the thread though if you want to read: http://duc.avid.com/showthread.php?t=8726

You are taking Nika's statement a little out of context in this particular instance. The potential problem that the dithered mixer was created to fix was due to repeated truncations. For example going from a series of 48bit plugs and truncating each time the signal was returned to the 24bit data path and also when the sizes of mixes exceeded a single dsp chip. In fact Nika did some tests himself and said the effect was noticeable (albeit under extreme conditions). Dither gets pretty complicated the more you look into it as in reality it's not really noise but a mathematical statistical construct (a Triangular Probability Density Function). So when you start doing a whole series of truncations it's possible that the statistical probability could be affected. It's also possible some correlations might occur but alternatively constantly summing multiple dither operations could have some audible effect on the level of the noise floor. Anyway, that was a discussion for a decade ago, I'm not sure if anybody is particularly bothered today but in case they are, they can try out the dithered mixer for themselves!

G

[daeron80]

Thanks! I love learning more about this stuff. I'll have to re-read that chapter and see if I can figure out where I went awry in my understanding.

[Greg Malcangi]

Your understanding isn't wrong under normal conditions but this wasn't a normal condition. The manual application of dither is a one shot use during mastering or, the dither applied during ADC or DAC reconstruction. But in this instance we are talking multiple successive truncation operations and summing the results!

The only part of what you said which was definitely wrong was that dither is noise that has been shaped. TDPF dither is not noise shaped and you would only use noise shaped dither in a final re-quantisation process (usually the final stage of mastering).

G

[daeron80]

Originally Posted by Greg Malcangi

The only part of what you said which was definitely wrong was that dither is noise that has been shaped. TDPF dither is not noise shaped and you would only use noise shaped dither in a final re-quantisation process (usually the final stage of mastering).

Gotcha. Or one could say TDPF is shaped flat. But so is hiss from, say, line amp transistors. Or a condenser mic circuit. Near enough, I would think.

BTW, that thread you linked to wasn't about dither but it was fascinating. I wish Lynn Fuston were still active here. I've sat in on his sessions. He really knows his stuff.

[Greg M]

You can probably talk to Lynn over on his 3DB forum if you are interested.

Greg

[daeron80]

Yeah, I check in there from time to time.

Mmmm. Well, I see was wrong about one thing: hiss from most sources has a Gaussian distribution probability which doesn't give equal chances to whether an error will be rounded up or down. So there's that.

Still, if it's loud enough - well above the level that 24-bit dither would be applied - then the quantization errors will only be correlated to the noise. I think. Could be wrong. But if Gaussian - even pink or brown Gaussian - is loud enough to completely bury any TDPF dither that were added, how is the dither going to do anything?

[Greg Malcangi]

Originally Posted by daeron80

BTW, that thread you linked to wasn't about dither but it was fascinating.

In a sense it was because we were discussing the internal routing of PTHD and what truncations were happening when not using the dithered mixer. If you liked that thread, get your teeth into this one: http://duc.avid.com/showthread.php?t=8396 and
http://duc.avid.com/showthread.php?t=5899 +
http://duc.avid.com/showthread.php?t=5549

Originally Posted by daeron80

Mmmm. Well, I see was wrong about one thing: hiss from most sources has a Gaussian distribution probability which doesn't give equal chances to whether an error will be rounded up or down. So there's that.

Still, if it's loud enough - well above the level that 24-bit dither would be applied - then the quantization errors will only be correlated to the noise. I think. Could be wrong. But if Gaussian - even pink or brown Gaussian - is loud enough to completely bury any TDPF dither that were added, how is the dither going to do anything?

It all starts getting a bit complicated. I have found that you get to a certain point in understanding the finer details of digital audio and to get any further you have to forget about thinking in terms of sound and approach it as a mathematician (yuck!). We want to think of dither in terms of white noise because that is what we hear at the end of the day but the reality of course is that dither is a mathematical equation. For example, as a sound guy we could ask the question "where exactly is dither applied, before or after truncation?" but the answer isn't easy to comprehend in terms of the way we usually process sound because the answer is "during". Dither is all based on probability (statistical) math, so messing around with "white noise" only creates a certain probability of it acting or not acting as dither. I agree that the result of the dither process (un-correlated white noise) can be masked effectively with many types of noise but the question is whether those other types of noise (say the noise generated by resistors in the analogue front end of the ADC process) has the correct distribution to create the complete randomisation of all quantisation errors that a TDPF is designed to? If it doesn't, then you don't have to worry about whether one type of noise masks another as it's a possibility that you may have strange harmonics in the program material many dB louder than the noise floor, due to correlated quantisation errors.

G

[daeron80]

That makes sense. I love math, actually. Thanks for the other thread links!