Topic: 32Bit Float recording - The Technical view

[datbrad]

From what I've learned, the generation of recorders that came out around 2017 use dual converters but instead of writing the lower level safety track, a circuit after the ADCs compares samples in real time and writes the ones with the best dynamic range, discarding the rest. I speculate 32bit float was just the next step in evolution by using the dual ADCs to write the larger range of 32bit float instead of having to keep within the scale of 24 bit. It also might be cheaper, which could explain all the budget devices that use 32bit float now.

[Ozpeter]

Just putting on record here for posterity in this 32 bit float discussion, that Tascam seem now to have sorted out the confusion within their own company as to (1) whether 32 bit float files can be created from a single ADC at all and (2) whether their latest launched devices do indeed have only one ADC, creating 32 bit float files that have no better dynamic range than 24 bit files.  They have deleted some of their previous comments on the YouTube channel and replaced them with a new comment which reads as follows -

"32-bit float point recording and Dual ADC technologies do not necessarily go hand in hand. Dual ADCs give a far superior dynamic range, sure. But with a single ADC, 32-bit float recording is still possible and the clip gain can still be adjusted as such without any degradation. The only real difference between one or two ADCs is the disparity in dynamic range."

So now it is clear as I and others suspected, that recent low cost "32 bit float" recorders are using a single ADC and the dynamic range enhancement gained from two ADC's is not being achieved.  Given the confusion within Tascam itself, consumers are going to be confused too.  Hmmm.

[datbrad]

Yak....It's becoming more attractive to go back to using cassette tapes. I wonder if my old D5s still work......

[morst]

Yak....It's becoming more attractive to go back to using cassette tapes. I wonder if my old D5s still work......

A lot of rock and roll barely has 50 dB dynamic range, cassettes are awesome for that! (says morst while listening to a 1990 tape on the Nak 582 > F3 > WAV @32f96)

[Ozpeter]

I have quickly made a YouTube video explaining this whole thing. (32 bit float audio from single ADCs not being mentioned by manufacturers).

Ironically I used the sound from the video camera, set to stereo, placed close to a noisy PC - hardly an advert for my skills with audio.  But I'm short of time.  That's my excuse, officer.

https://youtu.be/pAyaWH-WX-M?si=RlfKDhlKs6eY35aZ

My channel is not monetised so I am not earning from that link.

[TheJez]

Thanks @Ozpeter and others for making us aware about the issue regarding single/multi ADC devices. With the risc of complicating things even more: I guess a device having two (or more) ADC's doesn't necessarily make it a very good recorder! Of course the quality of the analog pre-amps matters a lot, but also something that didn't got much attention at all here: The quality of the 'combining algorithm'! There are a zillion ways to do it, all with their pro's and con's. Even 'good' algorithms may work well most of the time but might introduce artifacts in certain specific situations. I'm not saying that they do, just saying that they might. And even if they do, it might well be unaudible. We simply don't know. What we do know is the patent by SD that describes a possible way to combine multiple ADC's into a single 32bfp recording. It doesn't even necessarily mean that SD actually uses the algorithms they describe in their patent. It probably does mean that other manufacturers will be using different ways to combine, simply to avoid patent infringement. (Or maybe they do use the SD ideas and won't tell, or maybe they are paying license fees to SD, we simply don't know...)
Bottom line:
- We now know for sure that a recorder that has 32bfp doesn't automatically also has multi-ADC.
- I guess that even if a recorder has multi-ADC, it doesn't mean it's a good recorder.

I'm almost wishing the manufacturers would have sticked to 24-bit main/safetytrack recording, so the users would still have control of the analog gain (to optimize S/N ratio, with extra safety margin due to the 2nd ADC) and the option to combine during post using the algorithm of their choice with the settings of their choice (to fix the loud clipping parts and/or improve S/N in quiet parts, whatever is needed).

[Ozpeter]

Yup, the converters are kind of at the end of the chain and what comes before is still (or more) vital as ever.  And yes, how those multiple converters combine their output can make a difference - some people claim to hear artifacts (but I don't).

You know, it seems to me that that there should be a law against describing any recorder as 32 bit float.  It should be described as single converter or dual converter or even quad converter.  The output format is dictated by the number of converters.  It is, of itself, not of great interest.  When the Tascam guy swore that you can't write 32 bit float audio from a single converter, actually he got the whole thing the wrong way round.  You can't write 24 bit audio from dual converters (well, not without throwing out data, which would render the dual converters pointless).  Maybe we should stop talking about 32 bit float recording here, and only talk about multi-converter recording.  Companies required to describe their recorders as dual or single converter devices wouldn't be able to pull the wool over consumets' eyes, like we're seeing recently.  Single or dual (multiple) converters make a difference.  Output formats don't.  They are simply a natural consequence of what comes earlier in the chain.  And that's how they should be described.

I feel another video coming on... 

[TheJez]

You know, it seems to me that that there should be a law against describing any recorder as 32 bit float.

Haha, I get what you're saying. But I don't have a general problem as such with recorders being described as '32 bit float'. I do think there is a benefit with 32bfp storage when recorders can do post-ADC stuff like equalizing or basically anything that could change the amplitude of the samples. Personally I just want a recorder that just records, but if post-ADC features (either in realtime or not) are present on a recorder (and people actually using it), then I'd be happy with 32bfp storage, regardless if it's a single or multi ADC device. It's just that a device being 32bfp doesn't say anything about the recording quality. Nor does single or multi-ADC. Unfortunately the manufactorers are abusing the 32bfp buzz to convince us into buying new equipment. Just like we had to buy 3D and curved TV's. (Remember those?) Hopefully manufactorers will become/remain clear and honest about what exactly they are trying to sell to us, so we can judge for ourselves if we need their stuff or not. And we need to be alert to not blindly fall for their marketing tricks!
Tonight I'll be recording my first loud gig with my new Tascam FR-AV2 and enjoying the dual-ADC 'no need to worry about the record level' feature. Looking forward to it! 

[Ozpeter]

Multi ADCs do allow for a wider dynamic range to be recorded.  But indeed, they don't help with anything before the ADC, except perhaps the designers will couple a dual ADC system with upstream components which can make that wider dynamic range worthwhile.  Well, possibly...

[TheJez]

Multi ADCs do allow for a wider dynamic range to be recorded.  But indeed, they don't help with anything before the ADC, except perhaps the designers will couple a dual ADC system with upstream components which can make that wider dynamic range worthwhile.  Well, possibly...

Yeah, or they can implement two crappy cheap analog signal paths feeding two crappy ADC's, combining them with a crappy algorithm and still wave the dual-ADC 32bfp flag to make it sell  .  But indeed, on the bright side: The multi-ADC technology, combined with 32bfp does open up options to further improve recording quality at a reasonable pricing, and fortunately some devices are actually doing this!

[Rairun]

You can't write 24 bit audio from dual converters (well, not without throwing out data, which would render the dual converters pointless).

You can in fact write 24-bit audio from dual converters! In my experience, with the playback system I use and the loudest listening volume I find comfortable at 0 dBFS, you can increase levels by 50-60 dB before quantization errors become just barely audible (you need to increase levels by ~118.5 dB to normalise quantization errors to 0 dBFS). Meaning: if you record a sound at the threshold of perception (defined by me here as -60 dBFS), you can normalise it to 0 dBFS, and the quantization errors will be insignificant - they will be raised to a level that would be barely perceptible in absolute silence (i.e. the errors are imperceptible when the quietest sound actually recorded has been normalised to 0 dBFS). This is to say 24-bit files have an incredible amount of dynamic range and give you a lot of leeway for editing!

In other words: manufacturers could easily use a dual ADC setup and then assign specific level ranges to write the results to a 24-bit file. Assign +4 dBu to 0 dBFS, write the data of the signal path with less gain to the -20 to 0 dBFS range, and the data from the path with more gain to -20 dBFS and under. You'd run into the microphone and preamp noise floors before bumping into quantization errors.

Don't get me wrong, at that point, I do find it better to write to 32-bit float! But you can absolutely see the benefits of dual ADC with 24-bit files, because this design is meant to address noise in the analogue signal chain, not digital noise.

[Ozpeter]

Interesting!  But is anything commercially available?

[Rairun]

Interesting!  But is anything commercially available?

I don't have one, but I believe the MixPre10 II engages all 3 ADCs when set to 24-bit. AFAIK you are free to set the gain (so you can clip digitally), but the multi-ADC design does its thing to reduce the analogue noise floor of the quiet passages.

[Rairun]

Someone correct me if I'm wrong, but I think Sound Devices had multi-ADC designs even before their products had a 32-bit float option at all.

[Gutbucket]

No reason a multi-ADC recorder can't be designed to save in whatever format is desired, including 16bit mp3 or whatever. Zoom F8N_Pro can be set to record 24bit fixed, and Zoom has confirmed that the recorder is using the switching dual ADC signal path when doing so. I believe the SD MixPre's can as well. 

As for quantization errors, that problem was solved decades ago.  Application of dither eliminates quantization errors.  Just apply it regardless of bit depth. It's an easy fix. Manufacturers apply dither when saving to 16bit, yet for some reason don't bother when saving to 24bit.  I supposed they decided back when that the recorded dynamic range of 24 bit was large enough that any quantization noise at such low levels didn't matter when folks were setting gain reasonably.  Or maybe they determined that there was already enough random noise in the signal noise at those levels to effectively self-dither.  If manufacturers still consider not applying dither at 24bits to be a legitimate design choice, while now simultaneously claiming that 32-bit float "solves the quantization problem", I smell rotten fish.

Again, whenever you see marketing materials use those stair-step drawings, best to stop reading right there because you are being lied to.

On the 'combining algorithms' TheJez mentions-
They are clever, they are cool, they are useful.  But philosophically what bugs me deep down about all this is that the quality of a digital recorder used to be defined by absolute linearity within it's bandwidth - being measurably bit perfect in/out within that range.  Use that available bandwidth anyway you like and be assured that whatever you put in you get back out again.  That includes sound-design folks recording ultrasonics and pitching them down into human hearing range, which is apparently where multi-ADC combining artifacts of some recorder were noticed.  This deeper problem is not about us not noticing or caring about the artifacts / non-linearities, it's about the shift from truly predictable linear transparent systems to ones which only need be perceptually transparent. I appreciate the useful extension of dynamic range.. when it is done right and effective.  But when adopting the tech required to do that sacrifices measurable, linear, predictability I don't like where it can be too easily taken.  It all too quickly becomes a marketing smoke and mirror game rather than one of measurable linearity.  Reminds me of the mp3 perceptual codec thing all over again in sacrificing true linearity for perceptual psychoacoustics.