Topic: 32Bit Float recording - The Technical view

[Rairun]

Just the quiet parts. It's not terrible, but I might try and fix it up.

I've just listened to your recording! It isn't the same issue described in the video. I hear the noise, but it's everywhere during the quiet parts, not just around the transients. It's either the microphone's or the preamp's noise floor, and I don't imagine you'd get a lot less noise with a different recorder unless your F3 is faulty. There could be other weird things going on too - I know for example that I get a lot more noise out of my Church Audio preamp if I put it down on a church pew or my desk. If I keep it on my body, it's fairly quiet (I imagine this has to do with the grounding).

[SMsound]

This idea is making waves on various recording forums/blogs I read:
https://www.youtube.com/watch?v=y3431uljZ2k

tl;dr, on a 32-bit recorder, the multiple ADC's kick in depending on how loud the source is. The problem is that the different ADC's have vastly different noise floors, so hit a loud transient and your recorder will choose the low-gain ADC, which adds a lot of noise right around the transient.

Now, because the noise is different/worse only around the loud transient, it's also extra hard to profile and remove with software (versus a consistent noise throughout the whole recording).

May not apply to you all. For me, I record sopranos who go into glass-breaking-mode sometimes after a gently piano intro, so this is indeed a problem.

[grawk]

so can you post an example of where it happened on one of your recordings?

also: https://taperssection.com/index.php?topic=192100.0;prev_next=next#new

[Ozpeter]

Having spent a happy hour reading about the subject (in documents dated prior to this video), it seems to be... complicated.  One writer suggested that indeed the noise involved in 32 bit float recordings is variable, not fixed, but it's variable in the area of -144dB so it is unlikely to be audible.  I'm too old to hear the noise apparently demonstrated in the video, but I did rather wonder whether the source being used was not a good choice.  It sounded like some kind of sound effects file with a lot of constant low frequency background noise, plus artificial reverb (but I could be wrong) - is it the reverb noise that's happening?  A much better test would be to simply use a tone generator, surely? 

Anyway, I would be much more concerned about distortion (clipping) than noise for real-world examples, and the whole concept of having no level controls and no meters appeals to me immensely.  The ability to walk into a classical chamber recital hall and position a 32 bit float recorder/mic at the sweet spot, turn it on and leave it to its own devices uncablednd unmonitored, and go to the bar for the performance, appeals to me immensely.  Would the client complain about the subtle variable noise floor?  I doubt it very much.  Audience coughing and phones ringing are likely to be a greater source of irritation.

But I do concede that ultimate perfection should be sought by ultimate perfectionists

[unidentified]

So has anyone heard this transient noise in practice in the field? 

[aaronji]

The voltage ranges for the ADCs overlap, at least in the Sound Devices implementation. That means that the sound right around the "switch" is likely to be in more than one post-ADC stream. SD's algorithm, based on regression models, enables rapid shifts from one ADC to another and since there is overlapping content, this can also be done very accurately. I would guess this can be fine-tuned to eliminate or, at least, minimize this problem. Maybe different in Zoom's implementation. I don't know and this is just speculation anyway...

[Rairun]

Having spent a happy hour reading about the subject (in documents dated prior to this video), it seems to be... complicated.  One writer suggested that indeed the noise involved in 32 bit float recordings is variable, not fixed, but it's variable in the area of -144dB so it is unlikely to be audible.  I'm too old to hear the noise apparently demonstrated in the video, but I did rather wonder whether the source being used was not a good choice.  It sounded like some kind of sound effects file with a lot of constant low frequency background noise, plus artificial reverb (but I could be wrong) - is it the reverb noise that's happening?  A much better test would be to simply use a tone generator, surely? 

Anyway, I would be much more concerned about distortion (clipping) than noise for real-world examples, and the whole concept of having no level controls and no meters appeals to me immensely.  The ability to walk into a classical chamber recital hall and position a 32 bit float recorder/mic at the sweet spot, turn it on and leave it to its own devices uncablednd unmonitored, and go to the bar for the performance, appeals to me immensely.  Would the client complain about the subtle variable noise floor?  I doubt it very much.  Audience coughing and phones ringing are likely to be a greater source of irritation.

But I do concede that ultimate perfection should be sought by ultimate perfectionists

Imho the question is: do you want audible noise around some transients (not all of them, just the ones that are too loud for the high-gain ADC, but a bit too quiet for the low-gain one), or do you want noise everywhere? Without multiple ADCs, if you're expecting the music to have louder passages, the one option you have is to be conservative and use low gain for everything, which means the quiet parts will have a higher relative noise floor. Sure, it's easier to profile the noise in post that way, but I'd much rather have less noise during the quiet parts in the first place. I could hear the noise around the transients, but I personally don't find it distracting at all.

[Gutbucket]

Foreshadowing back when..

This thread, dedicated to 32bit Float recording in itself but not addressing the details of how it is implemented in specific recorders won't really be of much practical interest to tapers.. other than being useful to dispel some academic misunderstandings about what it can and can't do.  In other words it will be mostly academic because what really matters is how its implemented in each specific recorder in question.

That's where the rubber meets the road and where all the current confusion lies!

jerryfreak, nice experiment. Thank you for it.

I don't agree with the last part of the final statement (in boldface); it even seems propagandistic to me--an attempt to make the shifting digital noise floor seem like a virtue when it isn't one. I don't mean that it's necessarily a defect, either, because if it's low enough at all times, no one will hear it shifting. But "it successfully evades detection" is the best that can be said about it if so.

--best regards

(Bolding above is my emphasis) Of course back then we were speculating on where potential issues could arise.  The current hullabaloo about noise-floor windowing artifacts during the hand-off between multiple ADC's is a specific aspect of this.   DSatz essentially called out the noise shift issue quite specifically.  Going back and reading his comment about successfully evading detection brings to mind a another thought on a somewhat deeper level..

There is a deeper fundamental change going on in this shift to multi-ADC designs, which is a shift away from a recording system that is fully agnostic/isotropic in respect to the recording of whatever content fits within its bandwidth, to one which gives that up in achieving some alternate aspect deemed more valuable for the intended use case, even though doing so introduces problems for less-common uses.  If evading detection of the ADC switching artifacts can be successfully arranged for the recordists who value not having to set levels far more than a truly isotropic data set, such a trade off is likely to be a welcome one for them - its good enough for their purposes by definition and they gain a quite welcome new feature, even if the scheme may not be good enough for outlier uses where a more fully isotropic data set is required for more esoteric applications such as recording ultrasonic signals, dramatically pitch shifting content, or whatever.

The machine has become specially optimized for certain uses, at the cost of being less optimized for others.  Of course many other features of any recorder are optimized for its intended use, but the digital recording itself previously has not been.. unless recording to a lossy format.

I'm reminded of the development of things like noise-shaped dither, and psychoacousticly tuned lossy data compression.  Those are useful tools, the successful implementation of which required careful determination of how much fidelity to the source is needed, prior to sacrificing what lies beyond the limits of perception.  Its an optimization for certain use cases, which makes for a tricky game in determining where those limits might be, how they might be different for different uses, and how close one is willing to get to them in seeking to leverage human hearing perception to advantage. Noise shaped dither moves the bulk of the noise to where it is perceptually less obvious. Lossy codecs minimize storage requirements in part by discarding data deemed perceptually irrelevant. The fundamental shift is from a complete data set that contains extraneous information to one that is perceptually equivalent yet not fully complete.  Fundamentally this is the same philosophical calculus based upon making a decision about what maters and what doesn't. If well implemented, it's not going to be s problem for most, but may for some.  Its a sacrifice of true fidelity for all use cases, for easier use for intended target use cases.

[SMsound]

Following up:
- How many ADC's do different recorders have (MixPre-3/6 ii, Zoom F3/F6, etc.)?

Presumably, one fix is to just add more ADC's to these little recorders, so instead of a 30dB noise jump on passages where it chooses the lo-gain ADC, you could have 5dB increments of background noise increase. I have heard that some 32bit recorders have 3 or more ADC's, but not sure which ones those would be.

[goodcooker]

The recording in that video that's getting folks talking about this is a recording of a water droplet.

Recordists making Foley or SFX recordings often use 60-70 dB gain to get those types of sounds loud enough to be usable.

I'm not surprised at all that after that type of gain adjustment that you are going to be able to hear system noise. Most people recording music or even dialog are not going to hear the low level interference from the ADC.

I'll keep staging my gain the old fashioned way using 24 bit.

[breakonthru]

Following up:
- How many ADC's do different recorders have (MixPre-3/6 ii, Zoom F3/F6, etc.)?

i believe the zooms have 2 and the the mixpres have 3, (according to mixpre patent filings)

[datbrad]

Curious, do any the Zooms (F3/F6) or SD Mixpre II's have the option to still record 24 bit instead of 32 bit float?

[adrianf74]

Curious, do any the Zooms (F3/F6) or SD Mixpre II's have the option to still record 24 bit instead of 32 bit float?

MP-6 II can record either 32 or 24.  F6 can record 32 or 24 or both and the F3 is 32-bit only.

[Gutbucket]

Following up:
- How many ADC's do different recorders have (MixPre-3/6 ii, Zoom F3/F6, etc.)?

i believe the zooms have 2 and the the mixpres have 3, (according to mixpre patent filings)

^ I think so.  The grist is how the switching between the multiple ADCs is handled.  Can it be managed in such a way that the inherent change in noise-floor at the switchover (which is unavoidable unless some sort of noise reduction is done ahead of the switchover) is made imperceptible.  To complicate things further I believe many multichannel ADC and DAC chips can be configured to combine their multiple internal units in parallel to incrementally increase dynamic range without switching between them.  That avoids the switching problems but only gains a couple dB for each additional unit in parallel.  Rather than switching between units set to target different levels, I assume that strategy does so by way of the signal being correlated across the multiple units while noise is not.  I think that's how the strategy works anyway.. this goes well beyond my knowledge comfort zone.  Perhaps both of these schemes are being used.

[Gutbucket]

Curious, do any the Zooms (F3/F6) or SD Mixpre II's have the option to still record 24 bit instead of 32 bit float?

MP-6 II can record either 32 or 24.  F6 can record 32 or 24 or both and the F3 is 32-bit only.

Keep in mind that the switching ADC architecture and the recorded file format are two separate, distinct things.

The relevant follow up question is: When the 32bit recorders that can do so are set to record 24bit files, does that mean only one ADC is used? Or is the machine still switching between ADCs in the same way as when in 32bitFP mode but writing a 24bit file?  I suspect the second might be the case.  If so it will not be a work around for the problem.