Is 32 bit float better than 24bit?

I actually agree with Roland here, and he has a good point. Common practice says we should avoid dithering as much as possible and keep files at their highest resolution until the end, which is the Mastering stage. Going from 32 bit into 24 bit “requires” dithering, as there is a reduction of bits where rounding errors are introduced. Whether you’ll hear a difference or not, that is where people differ in opinion. I personally like to have the peace of mind that I am working at the highest resolution all the way till the end. YMMV.

We had an interesting discussion about this a while back and the conclusion was that the UV22 offers shaped dither that should only be applied once at the very final stage. Although dither should be applied whenever the word length is changed it should be non shaped noise if you know that further bit reduction and dither is going to be applied at a later stage.

Split is right.

Any type of noise shaped dithering (i.e. UV22, POWr-3, MBIT+, etc) should only be applied as the very last step in the Mastering chain. But non-noiseshaped dithering (i.e Triangular) should always be applied any time bit reduction occurs. Seems like you and I agree on that.

That’s why I was saying that some people disagree with this and they find that, even at 24 bit, the resolution is high enough to void the merits of using non-noiseshaped dithering. This maybe true in some cases, like when working with a single (or few) stereo files. However, it could become a problem with a greater number of files. I think that dithering is a lesser evil than the effects of rounding errors due to truncation, and it actually increases the dynamic range. However, this doesn’t mean we should be care-free about applying it, since any type of processing degrades the quality of the files. The less we process, the better the quality.

Take care!

Work at 32 bit internally at all stages. Only dither down to 24bit and 16 bit at the mastering stage. I deliver mixes to mastering engineers in 32bit files. It just means with plug ins processing and all that, there is nothing to think about. You have virtually unlimited resolution for all practical purposes.

I would like to challenge anyone to have blind A/B test on hearing 0.0000126% distortion (rounding errors of 24-bit processing).

And I also have to remind you that 32-bit floating point doesn’t have more precision than 23-bit fixed-point (when using IEEE floating-point standard with 23-bit mantissa). It’s just gives you the precision within extreamly wide gain-range.

Brings to mind the test a while back where they connected Monster cables to speakers and then wire coat hangers. No one could tell the difference. :laughing:

:laughing:

Jarno,

I’m sure you’re right, but why not use dithering when it actually increases resolution and hence gives you a better quality file (even if you and I can’t hear it)? On the other hand, I always try to avoid extra processing (dithering in this case) if I can help it and keep the highest resolution until the end. Then I don’t have to worry about any of this. But if for whatever reason I have to lower the resolution at some point, then it gives me peace of mind to dither. If that doesn’t bother you, then good for you :slight_smile:

Take care!

You can very well use dithering when converting 32-bit-floating-point signal to 24-bit. It doesn’t hurt, but I find it pointless:

  1. Recording: Your converter doesn’t provide full 24 bits of significant information which means dithering (which only operates on that last 24:th bit) is meaningless.
  2. Processing single recorded 24-bit file: if your processing is not increasing dynamic range (expanding), same as above applies. The last few bits of the signal are pure rubbish … dithering is again meaningless.
  3. Mixdown: If you record your mixdown in decent level (peaking more than say -12dBfs) and the dynamic range of your mixdown isn’t ridicilous (far more than any natural sound source) any quantisation errors to 24bit are meaningless.
  4. Bouncing a pure digital source (modelling VSTi): Now you have no rubbish generated by your A/D converters, so you may once again theoretically get better results with dithering… but once again, in practise only if you have ridicilous dynamic range.

Now, when re-quantisising (note the word I’m using: re-quantisising … some uneducated people use the word dithering when they actually talk about quantisising) into 16 bits, I’m fully supporting dithering. When going to 16 bits the quantisation error may create distortion as high as 0.003% (compared to full-scale, up to 0.03% - 0.1% on quieter parts of the music). These numbers are in par with distortion of analog recording systems and this is BAD distortion, not the nice harmonic one. So when going to 16 bits: DITHER! In 24-bits deficiencies of your A/D conververters outweight the quatisation distortion and it’s up to you if you want to dither or not.

In practise I don’t bother to dither to 24 bits, unless I have very good reaon to do it, but always dither when going to 16 bits.

I must admit I tend not to dither when going 32fp>24 as I’ve never noticed any difference!

And nether has any (PT?) studio I’ve sent 24bit files to :smiley:

But as I almost always work 32fp, I do export mixes at 32fp and I do have cubase set to 32fp as I tend to do quite a lot of offline processing!

Of course shaped dither to 16bit final mix, usually with the UV22.

Jarno,

I already know all of that. I was simply stating that I choose to apply dithering in the rare cases when I need to reduce the wordlength (whether I hear it or not). It’s a personal choice :wink: In Sonar I was able to choose different bit depths for recording, mixing, bouncing and exporting. Since Cubase doesn’t have as much flexibility in this aspect (without me having to manually change settings), then I just set it to 32 bit and forget about it (even if my converters don’t go near that far).

BTW, the highest quality converters I know of go up to ~21bit (or -127dB FS). These are the Lavry Gold. I personally don’t know of converters that go lower than that, as you mentioned in an earlier post above. Perhaps you know of ones that do? If so, let me know.

Take care!

I think we agree with principles but disagree with our choises. Neither us shouldn’t be right or wrong. It’s just matter of choise.

No I don’t. I haven’t bothered to search for those last extra bits. I don’t record classical music and I don’t think any of my mic/pre combinations have low enough noise floor to take advantage of more than 19 to 20 bits of A/D conversion. My original quote of “21-23 bits” was just for being safe: If I had said 21 bits, I bet there would have been someone with converter capable of producing 22-bit A/D and I would have been hammered down completely.

Not at Xmas… Surely

Maybe you’re right. We aren’t at GearSnobs.com after all :stuck_out_tongue:

Thank goodness for that :laughing:

Amen, bro! … while still have to question your choise to save your files as 32bit, because according to science it should be pointless unless very very very rare situations … but that’s again only the choise you make. And I can live with it … and I hope you too can live with my choise…

Absolutely. My choice to save mixes at 32bit is basically because most of my mixes are done ITB and Cubase outputs 32fp so why go 24 then into wavelab and back to 32fp then down to 16 or whatever.

As I said earlier if I’m sending files out, I go to 24 and usually don’t dither.

Exactly :slight_smile:

No I don’t. I haven’t bothered to search for those last extra bits. I don’t record classical music and I don’t think any of my mic/pre combinations have low enough noise floor to take advantage of more than 19 to 20 bits of A/D conversion. My original quote of “21-23 bits” was just for being safe: If I had said 21 bits, I bet there would have been someone with converter capable of producing 22-bit A/D and I would have been hammered down completely.

Oh, I see. I don’t think you would’ve been hammered in this forum for saying that (at least not from what I can tell in the short time I’ve been here). I was simply curious to know if there were such converters that could go lower, cause that would be quite the feat.

Take care!

The advantage of recording in 32 instead of 24 bits is that one of the “native” word sizes of the x86/x64 is 32 bits. You’d have to convert your 24 bit values to 32 bit before doing any math on them. So from a programming perspective it makes perfectly sense.

As was stated earlier, there is absolutely no advantage in recording at 32 bits since the file simply gets padded with zeros at the end (AD/DA converters only go up to 21 bits MAX, the three remaining bits are filled with self-noise from the components. That’s a $4,000 converter btw). The only reason I use 32 bit for recording, me personally, is because it is simply easier to set and forget than to have to remember to switch Cubase back from 24 bit recording into 32 bit bouncing/exporting mode every time I work in a project.

Cubase does it’s internal processing at 32 bit resolution, no matter what bit depth the audio was recorded with. There is no “physical” conversion of the audio file taking place in Cubase until the moment you export, bounce or freeze a track. IOW, the computational math resolution of a piece of software has nothing to do with the resolution of an audio file.