32 bits floating point to 24 bits fixed point

[16/44.1]

Do i need to dither a 32 bits floating point to 24 bits fixed point?
Or can i save the 32 bits floating point to 24 bits fixed point without rounding off with dither?
Is there a need for it?
I don't think so, because that 24 bits LSB will be drowned in the analogue noise........right?
Some say that the noise is at atomic level at 20/24 bits, then there's no need for dither.
Who can clear that out?

[Greg B]

dithering when converting from 32 float to 24 bit won't hurt anything, but it can help just a little. I usually use powr3 dither whenever dropping bits.

[cemski]

sorry to ask, but i have to put a "why" after each of your sentences. :)

Originally Posted by Greg B

dithering when converting from 32 float to 24 bit won't hurt anything, but it can help just a little. I usually use powr3 dither whenever dropping bits.

[jupiter8]

http://www.cadenzarecording.com/imag...tingdither.pdf

According to this PDF there's no use dithering floating point sources at all.
I have only skimmed thru it so i'm not 100% sure why but this is one of the explanations. Basically the noise will not be 100% random. It will to some extent depend on the signal being dithered.

The sheer nature of a floating-point system negates the ability to add dither that is at a
constant amplitude and thus uncorrelated to the amplitude of the signal and of the same
bit depth as all bits to be truncated.

[Greg B]

"dithering when converting from 32 float to 24 bit won't hurt anything, but it can help just a little. I usually use powr3 dither whenever dropping bits."

Originally Posted by cemski

sorry to ask, but i have to put a "why" after each of your sentences. :)

I said what I said based on personal experience, and from what I've read in the many other discussions on this topic, including what Bob Katz said in the book I'm reading in my picture.

I think it's good to try different dithers, and decide which you like best on a particular mix. What sounds best on one mix doesn't always sound best on another. There are subtle differences.

If you're not sure if dither sounds ok coming from 32-bit float, then try it. Do what you think sounds best. If you can't hear a difference, then don't worry about it.

[Lupo]

Originally Posted by Paul Frindle

Hmm.. Im not at all impressed with this much advertised 'dither difficulty' in float systems. I am not sure where this idea comes from and why people believe it. It really isn't true in practice..

To understand all this (quite simply with common sense) it is only necessary to think of float and fixed point math conventions as nothing more than numerical representations. Providing that the data is presented at the required precision with the required dither for the eventual target output data width, it is quite irrelevant which particular number system happens to be used for the processing itself. They are just numbers - and there is no magic :-)

So accordingly, if a float processing system must finally present data to a fixed point media (such as CD - or anything else in the real world) all that is needed is to add the dither at the correct level within the float domain as the final process, so that when it gets transfered to the final output data media the dither is correct for that output data width. Providing the float system can calculate to an equal or greater accuracy than the output fixed data width, there is nothing preventing the dither from working identically - regardless of whether it was generated in float or fixed math representation.

Quote lifted from this thread.

[Boomshanka]

Originally Posted by jupiter8

http://www.cadenzarecording.com/imag...tingdither.pdf

According to this PDF there's no use dithering floating point sources at all.
I have only skimmed thru it so i'm not 100% sure why but this is one of the explanations. Basically the noise will not be 100% random. It will to some extent depend on the signal being dithered.

True whilst within the realm of 32 bit fp but you gotta leave that world to go 24 bit fixed AES or 16 bit to CD. SADiE is one example workstation running 32 bit float. No need to dither internally but certainly whenever leaving the DAW or going to destination media.
So, generally best to dither to minimize your losses.

Originally Posted by 16/44.1

Some say that the noise is at atomic level at 20/24 bits, then there's no need for dither.

Certainly, cumulative 20bit truncation can be audible depending on the material.

[Bob Olhsson]

When you don't dither you are always adding unnecessary distortion. There have been many rationalizations for software and hardware to not dither however it remains theoretically necessary so it's pretty silly to not do it when the capability is available.

[tamasdragon]

You can try both, although I guess (depends on the material), dithered will sound better. My vote is to dither.

Tamas Dragon

[dkatz42]

It's worth noting that going from 32 float to 24 fixed is a lossy conversion unless the material is running at 0 dBFS. Some number of low-order bits are going to disappear in the conversion. Whether there is any meaningful data in those low order bits depends on how the material was processed, but in general this conversion is quite analogous to going from 24 fixed to 16 fixed, and so it would seem that dither would be a good thing.

On the other hand, if the final product is going to be 16 bit, there's nothing to be gained in dithering the first conversion, since any difference caused by the dither is going to be lost in the conversion to 16 bit.

[jupiter8]

Originally Posted by Boomshanka

True whilst within the realm of 32 bit fp but you gotta leave that world to go 24 bit fixed AES or 16 bit to CD. SADiE is one example workstation running 32 bit float. No need to dither internally but certainly whenever leaving the DAW or going to destination media.
So, generally best to dither to minimize your losses.

Well,i DO know that dithering isn't necessary when if you're not changing the bitdepth. Not sure what you mean by that.

However the PDF gives some convincing arguments that it is no use dithering when going from a float source to an integer target.
As i said i only skimmed thru it. I thought it'd be interesting to read an opposite opinion.

I guess this is one of the cases where you simply have to read a lot of opposite opinions and simply decide which one you believe in. Could be wrong about that though.

[tamasdragon]

It's an endless debate whether floating or fixed point is better. It's not a win win, but rather a situation of compromise. One is better at one thing, the other is better in others. To understand the whole picture you should look after some dsp engineering.
All in all, both can give you excellent results.

Tamas Dragon

[jupiter8]

Originally Posted by tamasdragon

It's an endless debate whether floating or fixed point is better. It's not a win win, but rather a situation of compromise. One is better at one thing, the other is better in others. To understand the whole picture you should look after some dsp engineering.
All in all, both can give you excellent results.

Tamas Dragon

You're absolutley right about that.But that's not really what we're discussing here,is it ?

[MattGray]

In this particular case going from 32bit float to 24bit integar or when going 48bit fixed point to 24bit, I've found that I get the most transparent result by using flat 24bit TPDF dither (no noise shaping). You should really only use 'noise shaped' dither on your final wordlength reduction to 16bits. It's the cummulative affect of doing multiple 'noise shaped' dither processes that can raise the noise floor to an unacceptable audible level.

Some tests I conducted with some mix engineer friends of mine involved using a Pro Tools LE system (32bit float). The output of the master in Pro Tools LE is reduced to 24bit integar. When no dither was used we noticed a slight smearing or fuzziness to the mix particularly if there was more processing to be done later i.e. mastering. Yet the same mixes with 24bit TPDF dither didn't exhibit the same fuzziness about them & provided a more transparent sounding master.

Yes this is only a very small improvement & only the trained ear would probably notice it, but it's something I now recommend that all mix engineers do for the best results. Ultimately it's up to you though, do a test & see if you hear the difference? To help you hear the difference add some compression or limiting to both samples after they've been converted to 24bits as this will help bring out anything ugly to the surface.

Matt

[jupiter8]

Originally Posted by MattGray

In this particular case going from 32bit float to 24bit integar or when going 48bit fixed point to 24bit, I've found that I get the most transparent result by using flat 24bit TPDF dither (no noise shaping). You should really only use 'noise shaped' dither on your final wordlength reduction to 16bits. It's the cummulative affect of doing multiple 'noise shaped' dither processes that can raise the noise floor to an unacceptable audible level.

Some tests I conducted with some mix engineer friends of mine involved using a Pro Tools LE system (32bit float). The output of the master in Pro Tools LE is reduced to 24bit integar. When no dither was used we noticed a slight smearing or fuzziness to the mix particularly if there was more processing to be done later i.e. mastering. Yet the same mixes with 24bit TPDF dither didn't exhibit the same fuzziness about them & provided a more transparent sounding master.

Yes this is only a very small improvement & only the trained ear would probably notice it, but it's something I now recommend that all mix engineers do for the best results. Ultimately it's up to you though, do a test & see if you hear the difference? To help you hear the difference add some compression or limiting to both samples after they've been converted to 24bits as this will help bring out anything ugly to the surface.

Matt

I'm not doubting your experience or anything and if it works for etc.
I'm partly here to share the things i know and to learn the things i don't know.

However 2 things.

1. If you have a stereo master at 32 bit float, would'nt the dither noise or quantiztion distorsion be well below the accuarcy of the DA conversion ?

The best figure i've seen is like 127 Db of dynamics range (The Lavry Blue) while the theoretical DR of a 24 bit file is 144.
So at least in theory you should'nt be able to hear the difference simply 'cause the converter isn't that accurate.

2. You say that repeated dithering has accumulative effects. True that but so has quantzition distorsion. And the reason for dithering is that it is harder to hear than distorsion. So i don't quite see your point there.

Did i miss something here ?

[masteringhouse]

I'm assuming that we're talking about a straight translation from 32 bit to 24 bit with no processing being done. This being the case, wouldn't whether or not to use dither depend on the actual level of the signal?

If you haven't gone beyond the dynamic range of 24 bit fixed, the 24 bit mantissa of a 32 float number would equate to 24 fixed, therefore it would seem that you are adding unnecessary noise to the signal by dithering. OTOH if it exceeds this range, as Nika's paper points out, you are adding correlated noise during the translation by using dither. Not really a good thing, but a necessary "evil".

[Bob Olhsson]

Dithering doesn't hide the distortion. It turns it into something that sounds like hiss.

Dither increases the noise level 3 dB. with every subsequent process.

Failing to dither creates distortion that increases by 6 dB with every subsequent process.

The bottom-line is that its a good idea to use dither if you have it but not the end of the world if you don't have the ability to dither to 24 bits.

[dkatz42]

Originally Posted by masteringhouse

If you haven't gone beyond the dynamic range of 24 bit fixed, the 24 bit mantissa of a 32 float number would equate to 24 fixed, therefore it would seem that you are adding unnecessary noise to the signal by dithering.

You're missing the fact that the float signal, while it only has 24 bits of precision, has more than 24 bits of dynamic range, and you lose data if the signal is less than 0dBFS. In other words, the problem is not that the signal is greater than 0dBFS (which will always clip), but rather that it is less than 0dBFS.

Simple example--start with an analog source that peaks at -6 dBFS and digitize it at 24 fixed. You have 23 significant bits (the high order bit is zero.) Convert to float; all of the precision is preserved (tack a zero into the exponent field and then normalize the number.) Pull the fader back by 12dB. At this point the mantissa still has 23 bits of valid data, but the exponent has decreased by 2. Now convert back to fixed. The mantissa gets shifted two bits to the right and then truncated; you're down to 21 bits of precision (and have three leading bits of zero rather than just one.)

My read of the cited paper is not that dithering is useless nor unnecessary on floating point format, but rather that it is impossible to fully decorrelate the quantization noise from the source signal, and thus may cause more harm than good if not done well.

[masteringhouse]

Originally Posted by dkatz42

Simple example--start with an analog source that peaks at -6 dBFS and digitize it at 24 fixed. You have 23 significant bits (the high order bit is zero.) Convert to float; all of the precision is preserved (tack a zero into the exponent field and then normalize the number.) Pull the fader back by 12dB. At this point the mantissa still has 23 bits of valid data, but the exponent has decreased by 2. Now convert back to fixed. The mantissa gets shifted two bits to the right and then truncated; you're down to 21 bits of precision (and have three leading bits of zero rather than just one.)

As I said in my inital post "I'm assuming that we're talking about a straight translation from 32 bit to 24 bit with no processing being done." That was my impression of the question by the OP, e.g. exporting an already existing 32 bit float signal and going to 24 bit.

As long as the mantisa of a floating point represents the same range (approx 6 db per bit intervals) the coded words would be translated to 24 bit more accurately without dither given correlated noise, true?

Another thread on this topic that reinforces many of the comments made on this thread:

click here

[dkatz42]

Originally Posted by masteringhouse

As I said in my inital post "I'm assuming that we're talking about a straight translation from 32 bit to 24 bit with no processing being done." That was my impression of the question by the OP, e.g. taking an already existing 32 bit float signal and going to 24 bit.

As long as the mantisa of a floating point represents the same range (approx 6 db per bit intervals) the coded words would be translated to 24 bit more accurately without dither, true?

If all you're doing is taking a signal in 24 fixed, converting it to 32 float, and then converting it back to 24 fixed, then it is true that this is lossless and dither will only corrupt the signal, since as you point out a 24 bit fixed value can be exactly represented as a 32 float (and thus converted back.) However, it's not terribly useful or interesting (you can accurately play back a single recorded track if absolutely nothing is done to that track other than the conversion back and forth.)

The real question is not whether processing is happening as part of the conversion to 24 bit (ultimately that is never the case) but what the source of the floating point data is in the first place.

In the real world there is always *something* done to the data (something as simple as a gain change, or as complex as you would like) between the initial conversion to floating point and the conversion back to fixed, and that causes issues because the opposite is not true--there are a very very large number of 32 bit float values that cannot be exactly represented as 24 bit fixed (on the order of 4.2 billion, or 99.2% of the possible values, if my math serves.) Thus, you end up with quantization and roundoff errors if there is any computation done from the point at which the original fixed-point data is converted to floating point.

[Spartacus]

Originally Posted by Bob Olhsson

Dithering doesn't hide the distortion. It turns it into something that sounds like hiss.

Originally Posted by masteringhouse

This being the case, wouldn't whether or not to use dither depend on the actual level of the signal?

If you haven't gone beyond the dynamic range of 24 bit fixed, the 24 bit mantissa of a 32 float number would equate to 24 fixed.

So, thing is, is there any distortion if we stay under a 24 bit dynamic range?
I mean, does for exemple a dance track using only the highest bits and never the ones under around 10 (except maybe for the fade in and fade out) will create any distortion?

[Spartacus]

Originally Posted by dkatz42

Thus, you end up with quantization and roundoff errors if there is any computation done from the point at which the original fixed-point data is converted to floating point.

But is this true for all level signal? or only the low level signals using the lowest bits?

i mean for this 24 bits word : abcd efgh ijkl mnop yyyy yyyy

Won't we always have for high level music the same y=1 ? and no truncation to take care of anyway?

i know lot of people will answer me "yes there will be truncation if some processing has been done"
but if the only process is to master an already high level signal (dance music,etc...), i would never touch the yyyy yyyy anyway right? and "y" will always keep the same value anyway as i will never put down the level so low to change it

[MattGray]

Originally Posted by jupiter8

1. If you have a stereo master at 32 bit float, would'nt the dither noise or quantiztion distorsion be well below the accuarcy of the DA conversion ?

Although the quantization distortion is very low down once you add further processing it brings to light the aliasing in a very small manner. As I said if you add compression or limiting to such a truncated signal you are bringing up the noise floor so you can hear it more clearly this is when it becomes most obvious.

2. You say that repeated dithering has accumulative effects. True that but so has quantzition distorsion. And the reason for dithering is that it is harder to hear than distorsion. So i don't quite see your point there.

Did i miss something here ?

Yes the fact that I was talking about 'Noise Shaped' dither. Because of it's aggressive nature you are more likely to hear multiple passes of 'Noise Shaped' dither if it's used repeatedly throughout your project. How soon it becomes audible depends on the nature of the source material & the amount & type of dither applied in succession.

This is why I recommend using TPDF dither 'no noise shaping' at 24bit level. Then apply one pass of Noise Shaped dither for your 24 to 16bit wordlength reduction. Hope this makes sense?

Matt

[masteringhouse]

Originally Posted by Spartacus

So, thing is, is there any distortion if we stay under a 24 bit dynamic range?
I mean, does for exemple a dance track using only the highest bits and never the ones under around 10 (except maybe for the fade in and fade out) will create any distortion?

I think that the following post from Nika sums this up nicely. Sorry for any confusion in my previous post, it had to do with quantization intervals matching between 24 fixed and floating point equating to each other. I did not express the question accurately.

See:
http://www.pgm.com/pipermail/proaudi...ry/002500.html

[Boomshanka]

Originally Posted by jupiter8

Well,i DO know that dithering isn't necessary when if you're not changing the bitdepth. Not sure what you mean by that.

I was responding to your "According to this PDF there's no use dithering floating point sources at all."

Originally Posted by jupiter8

I guess this is one of the cases where you simply have to read a lot of opposite opinions and simply decide which one you believe in. Could be wrong about that though.

There's ultimately a subjective trade-off between truncation distortion and dither noise (hence the many options and noise shapes available) but theoretically - and simply put - it's best to dither whenever reducing wordlength, be it to 24bit or less.

And for any confused, keep in mind we're talking dynamic range in terms of easily cumulative low level loss of detail which can affect sense of depth and "analog-ness" - it's "footroom" (to coin a BK phrase), rather than 0dBFS headroom.

[Pixel25]

so now I'm a little confused.
If I capture my mastered audio into my workstation via an L2,
which dithers it from 24 -16, do I need to switch on the dithering
in my DAW whe I burn a CD?
What about if there's no fades or edits?

[Pixel25]

another question;

Is it ok to use a 32 bit plugin on a 24bit audio file?
Am I doing any damage?

[MattGray]

Originally Posted by Pixel25

so now I'm a little confused.
If I capture my mastered audio into my workstation via an L2,
which dithers it from 24 -16, do I need to switch on the dithering
in my DAW whe I burn a CD?
What about if there's no fades or edits?

If there is no further gain changes, processing or fades to be applied after your L2 dither then there is no need to do any further dithering. Remember this general rule & you should be alright - You should always dither when there are any gain changes or wordlength reduction. Use 24bit TPDF dither when you are making gain changes with 24bit files & use 16bit noise shaped dither as the last step before burning to CD.

Matt