ORIGINAL: pattor

I have one more question - if a floating point "mix" would/could be altered into a integer word with 64 bits of resolution and then dithered to either 24 or 16 bits (in a 64 bit integer environment), would that method be a source for a more effective dithering, and - is it even possible, technically, to create such a scenario?

Absolutely, and of course!

Nika

"Absolutely" = it is possible to do this? Or = it would be if it was possible?

It is absolutely possible to do this.

"Of course" = dithering would take care of quantization distorsion completely?

No, as I stated above, in a floating point system you never completely remove quantization distortion - you only reduce it below the threshold of audibility. As I also stated above, I believe that in a properly implemented 64 bit system you would reduce that distortion to below the threshold of hearing for any practical situation.

So...is that scenario created in any floating point software of today?

I certainly believe so? But I'm not an expert on what is happening these days in software design. I thought Nuendo was doing this?

Nika

ORIGINAL: attalus

Nika, do you have a website where i can read more info from you on this topic later?

www.cadenzarecording.com for starters. If you're really brave there's also this site:

http://www.amazon.com/gp/product/141960001X/ref=sr_11_1/102-6912443-0376969?%5Fencoding=UTF8

Nika

ORIGINAL: Nika

ORIGINAL: pattor

I have one more question - if a floating point "mix" would/could be altered into a integer word with 64 bits of resolution and then dithered to either 24 or 16 bits (in a 64 bit integer environment), would that method be a source for a more effective dithering, and - is it even possible, technically, to create such a scenario?

Absolutely, and of course!

Nika

So basically your saying when dithering in a 64bit fixed point enviroment you'll come out with a cleaner sound then 64bit floating point?

Andy,

Of course I remember you. You were on the peninsula when we first talked and then moved to CA if I remember correctly. Also, if I remember correctly you're first name isn't really Andy?

And for the record, I'm back in Seattle now.

I'm glad to hear all is well with you.

Cheers!
Nika

ORIGINAL: Nika

So basically your saying when dithering in a 64bit fixed point enviroment you'll come out with a cleaner sound then 64bit floating point?

Well, now, I didn't say that, exactly. But if fixed point and floating point systems are both done properly, the distortion due to quantization error will be lower on the fixed point than on the floating point.

But floating point may benefit the sound more in other areas, right?

Nika

Ok, thanks! We had a discussion here awhile back on the differences of 64bit float vs 64bit fixed so i'm always curious on this (but a layman).

ORIGINAL: cGar

Hi Nika,

I just followed the link to your book and was considering buying it. But I had a few questios. (sorry all as this isn't totally relevant to the thread). Are you familiar with Ken Pohlman's Digital Audio book? That what I have along with sever DSP texts. I was wondering if there is complete overlap between yours and his or is there alot of new info that I could take away.

If you want to email me directly its christophergarrido@gmail.com

thanx,
chris.

Chris,

I have a copy of Pohlmans' "Advanced Digital Audio" here and I read it thoroughly when I was immersed in the topic. I also have Rabiner and Gold's "Theory and Application of Digital Signal Processing, Bob Katz's "Mastering Audio" and somewhere on the shelf here I have Watkinson's book.

If I thought I was merely creating an overlap I would not have written the book. It took about two years to complete the project and there's no way on earth it will ever pay off. I was under no delusions that it was not a 'for profit' venture, but I'm not a masochist, either. The reason I put it out there was because I thought that the information was lacking in the market - at least in a coherent, intelligible format. Apparently many others have agreed, including those whose quotes are on the back cover and the others who participated in the project.

If you have any other specific questions let me know. I hope this post is not out of place on this forum.

Nika

book@cadenzarecording.com

ORIGINAL: UnderTow


But you can allready! (At least since Sonar 5.2 and also with previous version if you run your soundcard at 16 bit).

UnderTow

i haven't used 5.2 yet .. but hope to soon. and .. again .. i don't dither . if i run a 32bit wave .. it's gonna
stay that way until the mastering engineer gets it .. it i run 24bit .. it stays 24 bit all the way through. i'm not
going to compound the shaping.

jeff

Thats a really good question!!!

According to the dither ALWAYS when truncating rule, the answer should be yes!!! One question is whether the powercore is already handling this on the way in or does sonar have to do it. Another is what do ALL the plugins that operate at double precision do when they return the processed samples back to Sonar at single precision??? They should be dithering right?

I would think that you would defintely want to dither from 48 to 24 when you leave a plugin? But what about 64-bit float to 32-bit float in a double precision float plugin. Probably yes too.

Anyone have the answers to any of these questions?

In theory TC should take care of any dithering needed. I have no idea if they do but they are a reputable company with good products so they probably do it right.

For other plugins, some do, some don't. You could test your plugins by sending a signal through them and see what comes out at the output.

As for dithering in floating point, I read this article about the subject a while back. Looking at it again now I see it was written by Nika. :)

http://www.cadenzarecording.com/images/floatingdither.pdf

UnderTow

What would be nice for you to compare would be your 4 consecutive dither example to 4 consecutive truncations!!! The RMS distortions may be slightly less but the peak distortions would be much higher and much more painful on the ear as each new truncation would only compound the previous distortions to other frequencies.

Thanks Cgar and undertow, when i have time i may do some testing

All the best

Matt

zungle...

When the recorded file enters any processing it will also enter a state of float. If the recorded file is 24 or 16 bits it is an integer file. In order to make sure what kind of format that Sonar read you can rightclick the audioclip and choose "properties". That will show what Sonar is reading from the first place.

As far as I know - Sonar will just read a recorded track which is situated on the hard drive. If you cut a recorded track so it's becoming shorter in Sonar, well what you do is just to tell Sonar to start reading a recorded track (situated somewhere on your hard drive) at some point and end the reading at some point, eventhough the original file might be longer.

But anyways, when processing occurs, the wordlength expand. The opposite of this is when the mix leaves Sonar to fit any integer wordlength. When the reduction of a wordlength occurs, the quantization distortion also occurs. This kind of distortion never happen when a wordlength expands or when Sonar is reading a integer file in order to put it into the mix.

What, on the other hand, will affect the original audio quality is many DSP (one DSP=one eq or one compressor or one reverb or one alternation of a track level or likewise) processings. Making the original audio go through a lot of these processings will maybe make the audio sound better in the big picture along with all the other tracks, but it will by each DSP process actually diminish some of the original audios "quality".

That is a normal tradeoff when working with mixes in the box, but might inspire people to maybe keep DSP processing to a minimum. As long as the plugin is having a great "manouvering room" by being 64 bits and Sonar is set to 64 bits and the plug send/receive at 64 bits, there should be no quantization.

When Sonar, on the other hand, will operate in 64 bits and send something to a 32 bit internal resolution plugin, to my knowledge there will be a slight reduction of the wordlength, hence also some quantization that will follow into the plugin. But, this quantization is also nothing much to worry about since those artfacts are thrown down at levels that are extremely low, so low that it does not matter at all.

But keep in mind that this scenario is based on that there has been one more DSP prior to sending to the 32 bit plugin. If Sonar send the 24 bit integer from the harddrive to the plugin (no trim is being applied within Sonar), I guess there will be no quantization at inpu because the wordlength in this case expand.

Maybe someone else know more about this issue.


EDIT:

When I'm trying to get a good overview of what's going on when routing in Sonar I kind of put up a mental picture of it. It's important to know the headroom (bitdepth) "inside" of the plugins and wheteher they send/receive 64 bit wordlengths in order to effectively "steer" audio into the best resolution and make the big mix movements there.

The mental picture of how audio flows and where it expands and shortens down within a mix software can make it easier to understand at which points audio might suffer more than at other points.

And to get rid of DSP processes that can be eliminated. One basic trick is to keep the master fader at unity gain and steer the mix to not clip the output. With the master fader at unity gain, you've eliminated one DSP that might be unnessecary.

If you send a single track to an aux return (say a reverb), you can send with the aux send at unity gain, have the plugin's output trim at unity gain and control the return with the return bus fader. If you instead adjust the send to, say - 4.3 dB, the plugins out gain to +4 dB and the return fader to - 12.3 dB, you've added two DSP processes that might be unnessecary.

EDIT_

ORIGINAL: Rednroll

Ok, So I did a little study to verify this. Here's what I did. I started out with a 16 bit silence waveform, 3 minutes in length and took an initial data point. Then I applied triangular dither and then did the same statistical data measurement. I then followed this by an additional triangular dither, get statistical data....rinse...repeat at each step, to see what the cumulative difference of what the dither could add on each step. What I found was like I mentioned before. Dither can add up to "+6dB". Maybe you can further explain where you got 3dB, because it's not what I'm seeing? This was all done within Sound Forge 8.0.

Thanks,
Red

Red,

Indeed, any time you sum two waveforms it is possible that the two peaks will add together at the same time, thus doubling the amplitude, thus increasing the peak by 6dB. The same is true with noise. If you add random noise dispersed between 1 and -1 together for long enough there is indeed a chance that at some instantaneous moment in time you will end up summing two 1s together or two -1s together and therein have a doubling of the peak amplitude - either 2 or -2.

But this isn't very reflective of how we hear, so this isn't how we discuss the amplitude of signals, nor the amplitude of noise. Looking at specific peak amplitudes is only really practical when looking at digital signal processing limitations. When trying to analyze what things sound like and the audible impact we discuss these matters in terms of RMS.

Now, if you take two identical sine waves and sum them together both the peaks AND the RMS values double (increase by 6dB), so it is easy to say that the signal doubles in amplitude. This is the rule when you have two "correlated" waveforms - in other words, when two waveforms are related to each other in some way then we say that summing them can increase RMS by "as much as 6dB." Indeed it will only be 6dB if the signals at RMS peak sum together at the same time - but this is not unlikely when the signals are related to, or correlated to, each other.

For non-correlated waveforms, though (for waveforms that are unrelated to each other whatever), when you sum the two together the RMS value increases by merely 3dB, despite momentary periods in time when the signal's peak value may indeed double.
Remember that we're interested here in what the ear hears and how. The ear basically hears a windowed RMS signal, not peak values.

Therefore, dither noise that is properly decorrelated from the signal itself only increases the signal's overall level (RMS) respectively. And dither added twice only yields an increase of 3dB. Dither added four times increases the dither amplitude within the result by 6dB, but remember that the overall amplitude of the signal also increases respectively! The result is that if you add two dithered signals together there can actually be an INCREASE in dynamic range - the signal increases by as much as 6dB and the dither noise within increases by 3dB, so the difference between the RMS values of each increases by as much as 3dB!

Let me know if you have any more questions. I hope this helps.

Nika

Since you are just putting the files into another digital editor (i.e. Audition) just export the files from Sonar as 32-bit floats. That way you don't have to dither when leaving Sonar. Then use the 32-bit option in Audition and only dither at the very end when you export to 24-bit or 16-bit from Audition.