32-bit float to 24-bit -- is this a problem?

[bobby yarrow]

I've been recording in SX3 in 32-bit floating point. Obviously, my a/d is 24-bit, so the 32-bit thing must be happening in cubase. I'm now using melodyne, which won't take 32-bit files. In SX3, I can quickly 'convert' the file to 24-bit -- seems instant, like no meaningful processing's involved.

I really think that I don't hear a difference. Are my ears deceiving me? Is this conversion, back to 24-bit, a delicate one, that should be addressed using R8brain or some sophisticated software, or am I safe just doing it quick and dirty?

Thanks.

[juicemaster1500]

To me freezing stuff to 24 bit files doesn't sound as good as the 'live' 32 bit source, when I started to add a L2 on tracks before freezing it seems to help (because of the dither I assume).

[genericperson]

i'm a bit junky myself. but 32 to 24 is not nearly as big of as crime as going from 24 to 16. at some point you have to get work done!

there are two main reasons you are not hearing the difference:
1) your playback system is probably average rather than world class
2) your converters (and everyone else's) max out at 24 bit d-a. so even if you are staying 32 bit, your converter or stereo buss is bumping it down to 24 bit at output.

it's a cruel world

[bobby yarrow]

Quite right, a cruel world. And, whatever the strengths of my monitors, consider the fact that I've played in and been around rock bands for about 20 years . . . I'm surprised I can still hear the doorbell, let alone 20K.

The thing with 32-bit is, I've never really gotten a handle on what it means to record at 32-bit fp thru convertors that produce 24-bit signal.

In cubase, you don't actually need to freeze or render a file at all to convert back to 24-bit. You just 'convert' it, and it's nearly instant. Think that's a destructive process, or just a restoration to what cubase got from the a/d originally?

[mogWai]

Originally Posted by bobby yarrow

Quite right, a cruel world. And, whatever the strengths of my monitors, consider the fact that I've played in and been around rock bands for about 20 years . . . I'm surprised I can still hear the doorbell, let alone 20K.

The thing with 32-bit is, I've never really gotten a handle on what it means to record at 32-bit fp thru convertors that produce 24-bit signal.

In cubase, you don't actually need to freeze or render a file at all to convert back to 24-bit. You just 'convert' it, and it's nearly instant. Think that's a destructive process, or just a restoration to what cubase got from the a/d originally?

Yea - there is no point in recording 32bit float files through a 24bit converter. Cubase will just add extra zeros and all you do is waste hard drive space.

Cubase does all its internal calculations at 32bit float anyway, and outputs a 32bit float audio stream - you really should then dither the output of cubase back to 24bit, for monitoring and to export a 24bit mix.

The reason you probably dont hear a difference is that we're talking about noise about 144dB down, and you've got to have some kind of super-sonic-sonar radar to hear things down there.

thats not to say its not accumulative though, and at some point those errors may manifest themselves as distortion in the range of hearing.

[David Lee]

couldn't you just stick dither (uv22hr) in line before the plugin to get it to 24 bit?

[BenF]

I experimented with recording at 16, 24 and 32 bit depth with 44.1/48/88.2 and 96khz sampling rates from tape. On my low end rig (RME DigiPAD and SparkXL) I found that the 32/88.1 sounded the most like the tape. But the tape still sounds better, of course. And this is a cassette (metal TDK on a Nak)! I recorded the material live of a choir in a nicely reverberant church.

The 32/88.2 retained more space and detail, and didn't change the sound as much. I know that mathematically the RME convertors can't do more than 24/96, so beats me why the difference. It doesn't make sense. And I wanted to record at less bits and sampling to save hard drive space. I was listening trying to make a case for 24/44.1 or anything that would save more hard drive space and sound the same. I couldn't do it. By the way, 96khz was worse than 88.1 so I'm guessing the clock in these cards was optimized for 44.1/88.2, rather than 48/96.

Now if I could just afford a new Weiss, Lavry, Mytek, or similar really capable convertor, I bet it would be different and I would like the result a lot more. I am disappointed with the conversion, but that's what low budget compromise is about.

I just thought sharing my experience may be of interest to someone.

Thanks. Ben

[bobby yarrow]

Originally Posted by David Lee

couldn't you just stick dither (uv22hr) in line before the plugin to get it to 24 bit?

Melodyne needs to load the whole sample, so I'd have to print the track to get the dither.

I'm really unsure about dithering to get from 32-bit fp to 24-bit anyhow.

As to why I use 32-bit at all, well, disk space doesn't bother me much, and I saw no reason not to do it. I don't record at a higher sample rate because of the huge hit on processing power, and because I'd have to upgrade 10 channels of a/d to do it on my present system.

[Reptil]

Originally Posted by BenF

But the tape still sounds better, of course. And this is a cassette (metal TDK on a Nak)! I recorded the material live of a choir in a nicely reverberant church.

on 30 on 45 GO!

creating 32 bit float files in Spark renders them unusable for any other application, including the finder (macincrash tiger panther). so it is 24 which is fine. Using 32 bit float for EXS24 (and I suspect other samplers as well - didn't test Kontakt yet) as a "storage" format, when sampling, does increase the speed of processing considerably. But this is a separate issue from HD recording.

[juicemaster1500]

Originally Posted by mogWai

The reason you probably dont hear a difference is that we're talking about noise about 144dB down, and you've got to have some kind of super-sonic-sonar radar to hear things down there.

No, the DITHER is the noise about 125-135 db's down, the effect of a dithered file as opposed to a truncated one is all over the place, in stereo imaging and less digital bullsh't artifacts.

[bobby yarrow]

So, um, given that I have 32-bit floating point files and I need 24-bit files, should I:

1. Just let cubase do the conversion (it's instant, not a mix-down or print or resampling);

2. Print the file at 24-bit, with dither; or,

3. Use some fancy software like r8brain to do the conversion.

Thanks.

[Lagerfeldt]

Just to make something clear:

1) Record your files in 24 bits.

2) Don't add dither when mixing or bouncing individual tracks, try to keep them in 24 bits or 32 bits as long as possible. Cumulativ dithering has a A) a tendency to veil the sound slightly, and B) can produce nasty artefacts when processed again (eq, etc.) due to the noise shaping algorithms (UV22, POW-r#2+3 or IDR)

3) Only dither when going from 32 or 24 to 16 bits, and don't touch the file afterwards at all, see B)

Furthermore, a 24 bit file recorded into Cubase for example, once changed even the slightest (eq, whatever) will expand the bitdepth to e.g. 32 bit. As someone else said, 32 bits to 24 bits isn't really a big issue, but going to 16 bits is.

[bobby yarrow]

Nice. Got it. Thanks.

[juicemaster1500]

Originally Posted by Lagerfeldt

As someone else said, 32 bits to 24 bits isn't really a big issue, but going to 16 bits is.

this is bullshit, if you can't hear it or hear it but don't care, ok.

no wonder a lot of people think digital sounds like ass

most amateur software developers don't even care.. no wonder their apps sound the way they do.. like shit

[dkatz42]

Originally Posted by juicemaster1500

this is bullshit, if you can't hear it or hear it but don't care, ok.

no wonder a lot of people think digital sounds like ass

most amateur software developers don't even care.. no wonder their apps sound the way they do.. like shit

Well...there basically is *no* difference in accuracy between 32 float and 24 fixed, since a 32 float has only 24 bits of data (and eight bits of scaling.) If you are storing the data coming off of a 24 bit interface, your data will be identical in either case (but 33% larger on disk if you store it as 32 float) as there are 24 bits of information, no more. The other eight bits will basically be the same on every sample (or else the samples will be bit shifted so that the high order bit is always set, and the zero bits that were on the top will end up on the bottom.)

Where 32 float is important is in mixing and processing. It increases the dynamic range to over 1500 dB, so you need not worry about interim bus levels and such (basically you cannot possibly clip, and the mix bus noise floor is at -1500dBFS, at least if the software is reasonable.)

For things like freeze files, 32 bit is helpful, since once you start processing those 24 data bits, it will keep you from losing low order bits (if, say, you have the gain down on a track) up until the point in the mix bus where the quiet tracks and loud tracks are added together. Effectively, you don't have to worry about the noise floor on those frozen tracks even if they are quiet, and you don't lose resolution. Imagine doing the equivalent with analog and you can see what would happen to quiet tracks.

There is no magic from going from 32 to 24, nor any inherent loss, so long as the levels are reasonable, and those lost bits were going to be lost anyhow at the D/A conversion. (If your mix peaks at -6dBFS, the low order bit will be tossed when the conversion is made to 24 fixed, since the high order bit is going to always be zero.)

Clear as mud, I'm sure.

[juicemaster1500]

yes I know a 32 bit and 24 bit file or 'flow' is supposed to be the same.

yet going from 32 to 24 bits gives audible artifacts, or loss of 'realness' in sound.

I don't like what I'm hearing when I do this, unless I apply some sort of dither.

[Absolute]

Originally Posted by juicemaster1500

this is bullshit, if you can't hear it or hear it but don't care, ok.

no wonder a lot of people think digital sounds like ass

most amateur software developers don't even care.. no wonder their apps sound the way they do.. like shit

we got Bleeder

[Lagerfeldt]

Originally Posted by juicemaster1500

this is bullshit, if you can't hear it or hear it but don't care, ok.

no wonder a lot of people think digital sounds like ass

most amateur software developers don't even care.. no wonder their apps sound the way they do.. like shit

First of all, check your attitude and language.

Then check your facts.

There's virtually no audible difference between 32 bit (float) and 24 bits (fixed), as in this example, but going from 24 to 16 bits is a much bigger issue. Multiple layers of dithering is absolutely not a good idea though and will degrade sound quality much more than truncating to 24 bits.

I believe dkatz42 explained the techincal side very aptly.

BTW, I just did a dithering test between the 3 kinds of POW-r and 6 kinds of IDR. No UV22 yet though. Interestingly, I only liked 2 out the 9 dithering algorithms, 2 were horrible to my ears. I'll post it later (it's a whopping 88Mb zip file).

[dkatz42]

Originally Posted by juicemaster1500

yes I know a 32 bit and 24 bit file or 'flow' is supposed to be the same.

yet going from 32 to 24 bits gives audible artifacts, or loss of 'realness' in sound.

I don't like what I'm hearing when I do this, unless I apply some sort of dither.

It's not clear what you mean by this, since by definition you are always listening at 24 fixed (since that's what your DAC uses.)

Can you explain a bit more about what you're comparing and how?

[bobby yarrow]

Originally Posted by Lagerfeldt

BTW, I just did a dithering test between the 3 kinds of POW-r and 6 kinds of IDR. No UV22 yet though. Interestingly, I only liked 2 out the 9 dithering algorithms, 2 were horrible to my ears. I'll post it later (it's a whopping 88Mb zip file).

First off, I appreciate that you're testing dither. It's just nice, that's all.

Wonder if the mbit dither in ozone is on your list. That's the one I use, over the options in SX3 and samplitude. Course, I'm not doing sample rate reduction, which might change things.

[cdog]

Originally Posted by juicemaster1500

this is bullshit, if you can't hear it or hear it but don't care, ok.

no wonder a lot of people think digital sounds like ass

most amateur software developers don't even care.. no wonder their apps sound the way they do.. like shit

Some people are so confused they have evolved advanced capacity to further confuse others.

[theblue1]

Originally Posted by dkatz42

Well...

... Clear as mud, I'm sure.

Thanks for the explanation! I realized I was unclear on 32 bit float and its use and implications in DAW software. That helps clarify a couple of points I was still kinda fuzzy on.

[GPl]

Originally Posted by Lagerfeldt

First of all, check your attitude and language.

Then check your facts.

There's virtually no audible difference between 32 bit (float) and 24 bits (fixed), as in this example, but going from 24 to 16 bits is a much bigger issue. Multiple layers of dithering is absolutely not a good idea though and will degrade sound quality much more than truncating to 24 bits.

I believe dkatz42 explained the techincal side very aptly.

BTW, I just did a dithering test between the 3 kinds of POW-r and 6 kinds of IDR. No UV22 yet though. Interestingly, I only liked 2 out the 9 dithering algorithms, 2 were horrible to my ears. I'll post it later (it's a whopping 88Mb zip file).

I would be interested in hearing this. Look forward to the post.

dkatz42 explained the technical side perfectly. Someone had to........

[juicemaster1500]

Originally Posted by Lagerfeldt

.

There's virtually no audible difference between 32 bit (float) and 24 bits (fixed), as in this example, but going from 24 to 16 bits is a much bigger issue. Multiple layers of dithering is absolutely not a good idea though and will degrade sound quality much more than truncating to 24 bits.

bullshit again,

recording a file as either 24 or 32 bits will sound basically the same, but GOING from 32 to 24 bits, the conversion or TRUNCATION, gives audible artifacts.

that's what I'm saying.

there's not necessarily a problem having multiple layers of dither, did you know that all the sony oxford plugins have buildt in dither (at least for tc powercore)?
there's no problems having multiple sony plugins after one another.

[juicemaster1500]

Originally Posted by dkatz42

It's not clear what you mean by this, since by definition you are always listening at 24 fixed (since that's what your DAC uses.)

yes the dac is 24 bits. that's what I'm actually listening to.

BUT, to get to the point where I can feed the dac a as good as possible signal, my signals goes between a lot of different bit depths. Sometimes I have a 16 bit sample, the daw software is 32 bit, a lot of plug-ins are 64 bit, some are 24 (!) bits etc.

to truncate from a high bit depth to a lower one is not a sonically pleasing consept.

dither helps smoothing out the process.

[dkatz42]

Originally Posted by juicemaster1500

recording a file as either 24 or 32 bits will sound basically the same, but GOING from 32 to 24 bits, the conversion or TRUNCATION, gives audible artifacts.

Going from 32 float to 24 fixed is not truncating; there are only 24 data bits in the 32 bit float. If the signal is at a lower level but actually has 24 bits of content (say, a fader is pulled down), low order bits will be lost in the conversion, but those bits would be thrown away at the DAC anyhow, which is a 24 fixed device. It's easy to demonstrate that you will end up with *exactly* the same bits going to the DAC in either case.

As I said earlier, the value in 32 bit floats are in dynamic range, not in resolution (other than a marginal improvement in roundoff error when doing lots of intermediate calculations.)

The other issue is that the bottom three or four bits in a 24 bit sample are garbage anyhow, as the converters simply aren't accurate enough to put something useful there. This means that dithering beyond 24 data bits, assuming that you had them (which you don't) is essentially fruitless.

The 24/16 conversion is a different animal, because it *is* very lossy, and dithering provides a way of preserving a couple of bits beyond the 16 in a statistical way.

[juicemaster1500]

yes I know, in theory it's possible it shouldn't matter, but it does anyway.

I think a part of the problem is that the programming monkeys that make daw software are using the wrong parts of the brain too much, and it has left them incapable to actually listen and determine what sounds good or not.

They lie to you and give you the very false impression that once a signal is 'in the box', it can't get degraded and it's all good.

In reality every single thing you do within a daw reduces the audio fidelity somehow, putting the faders to low, putting the faders to high, operating at 44,1k etc etc.

and yes, not adding dither when going from 32 to 24 bits.

[Lagerfeldt]

Originally Posted by juicemaster1500

I think a part of the problem is that the programming monkeys... They lie to you and give you the very false impression that once a signal is 'in the box', it can't get degraded and it's all good.

Hey, don't forget to take those anti-paranoia pills. :-)

My test only includes POW-r and IDR dithering for now. It can be found elsewhere on this forum (called The Dithering Test).

[Solari]

there's not necessarily a problem having multiple layers of dither, did you know that all the sony oxford plugins have buildt in dither (at least for tc powercore)?

Is it really like that? I mean, do the Sonnox Plugins really have an internal dithering for the case that a 32bitFP signal comes in? Even the native versions? So one doesn't have to worry about truncation here?

And what about TC Powercore in general, which is working with 24bit. Do i have to insert a dithering plugin in front of a powercore plugin (e.g. VSS3)?

I'm working with Cubase SX3 and TC Powercore MKII (predominantely VSS3)...And i take into consideration to buy the native Sonnox EQ, TransMod and Inflator...

Thanks,
Solari