Topic: Recording in 24-Bit vs. 16-Bit

[WiFiJeff]

Recent 32-bit experience, and why I don't intend to go back to 24 bit:

After a nearly 4 year hiatus since the onset of Covid, I went back last week to recording a weekly chamber music series that has just started up again.  I had not used any of my higher-end equipment since early 2020 and so knew I had to do some maintenance.  I replaced lots of batteries and power bricks that had died or lost reliability, and went with my Zoom F6 with triple-redundancy powering (and extras on the side).  My cables and stands were stored in a closet at the venue, so I brought a fresh set of XLR cables and some cable cleaning fluid.  I relearned how to run the F6 (a little nervous not to have the redundant second SDHC card that the F8n has, but with a loose extra card along anyway) and set it for 32-recording alone, no 24 bit parallel so I wouldn't have to worry about levels.  This immediately saved me from a heart attack when I found that the headphones stored at the venue were 1/4" plug but the F6 only has 1/8" output.  Yak!  But, hey, with 32 bit I could actually fly blind, and I had recorded enough piano solo in that hall to know how to setup without soundchecks, though I was not exactly happy about that (I arrived early enough that the lights in the room were not on, and I got through half my setup, from memory, in very very dim ambient  illumination).  I eventually found a 1/4" female to 1/8" male connector (and a second one turned up hours later), so could monitor the session.  I was recording two stereo pairs: two omni mics on a Jecklin Disk over the piano tail and a new set of wide cardioids (DIN, sorta) centered.  I monitored the Jecklin Disk set which was my usual main recording, the wide cardioids (Nevaton 59W-S) were experimental.  I bought them just before Covid and had never used them in the field.  The experiment was interesting, mainly I think because I did a bad thing.  In seeking to make sure a signal was coming from the 59W mics on the F6 meters, I unplugged and replugged the Nevaton mics _without turning off phantom power_.  Naughty naughty.  All four tracks were recorded as individual stems, not stereo linked.  When I got home and looked at the files in Wavelab, the omni tracks were perfect but the 59W had the right track at about the same raw levels as the omnis but the left almost invisible.  Quick loudness check in the wavelab "Analysis" menu showed the left track was almost exactly 20 dB lower (the 59W mics do not have pads, and tests of the mics on a 24 bit Tascam recorder showed the mics were pretty closely matched; the F6 settings for both channels were identical).  So I boosted the left track 20 dB, and ended up with a perfect stereo pair, no sign of any artifacts from system or background noise, which is exactly what 32-bit is supposed to do.  I suspect that fiddling with the connections during setup, with phantom on, caused a transient that "fooled" the F6 A/D to record the left channel lower.  Do the 32-bit A/D converters work in 20 dB jumps? Is this scenario possible or plausible?

Other things I forgot I needed: velcro cable ties, tape scissors for cutting tape to 1) replace velcro cable ties and 2) narrow the Shapeways printed Nevaton 59-S DIN mount so it is actually the right size to fit 59-S mics.

Both recordings came out great, in my opinion.

Jeff

[Gutbucket]

The original versions of the MixPre (such as the MixPre-6) have multiple ADCs and record a maximum of 24-bit. They even cite their patent in the specifications in the user manual. This is also the case for the higher-end Sound Devices recorders, such as the Scorpio, 888 and 883.

Thanks. I was unaware of previous SD products incorporating multiple ADCs limited to 24bit. That just makes me wonder even more about all this.

This was discussed quite a bit previously; with those recorders, it is generally possible to set the gain such that the entire dynamic range of most microphones, which far exceeds the dynamic range of a performance, can be recorded.

Yes, and the question remains..  If the real-world dynamic range capabilities of the recorder is smaller than that of either file format, its the recorder's dynamic range limitation that is the bottle neck, not the file format.  In such a case, what is the need or value of switching to the 32bit fp file format? 

I see claims of post processing, but that's typically done in a 32bit fp workspace regardless of the bitdepth of the source files.

[aaronji]

Thanks. I was unaware of previous SD products incorporating multiple ADCs limited to 24bit. That just makes me wonder even more about all this.

Not just previous. The Scorpio is their current flagship model.

Yes, and the question remains..  If 24bits exceeds the analog dynamic range of which the recorder itself is capable, the recorder's dynamic range limitations that are the bottle neck, not the file format.  And in that case, what is the need or value of switching to the 32bit fp file format?  The real-world dynamic range capabilities of the recorder fits within either file format.

I see claims of post processing, but that's typically done in a 32bit fp workspace regardless of the bitdepth of the source files.

The need/value is you don't have to figure out the range of gain settings that work best with your mics. I think the appeal of "string everything together and press record" reigns supreme for a lot of us, for obvious reasons.

When I got my first gen MixPre-6, I recorded a show where the opening band didn't provide a board feed, but the headliner did. Somehow, though, a low-level signal was coming through. I didn't even notice until I got home. It was peaking at -33 or -35 dB, but there was zero noise when normalized to -1 dB. After that, I played around with it a bit and realized that I could figure out the gain setting such that, given the mic sensitivity and SPL handling specs, the input and the mic would simultaneously overload. Normally, I add a bit of gain to that setting (peak SPL is extremely unlikely at most shows I attend) and turn on the MixPre's excellent analog limiters. To my knowledge, the limiters have never actually kicked in.

[Gutbucket]

That the Scorpio is their flagship recorder targeted at professional uses, and does not include 32bit fp could be a confirmation my suspicion. 

The need/value is you don't have to figure out the range of gain settings that work best with your mics. I think the appeal of "string everything together and press record" reigns supreme for a lot of us, for obvious reasons.

I totally get that. But you should be able to "string everything together and press record" using a 24bit recorder that features the same multiple ranging ADCs in exactly the same way as with a 32bit fp recorder because the actual dynamic range thruput of the 32bit fp recorder itself isn't any larger than the 24bit one.  The increase in dynamic range that was necessary to accommodate that kind of flexibility is being provided by the multiple ADCs. 

When I got my first gen MixPre-6, I recorded a show where the opening band didn't provide a board feed, but the headliner did. Somehow, though, a low-level signal was coming through. I didn't even notice until I got home. It was peaking at -33 or -35 dB, but there was zero noise when normalized to -1 dB. After that, I played around with it a bit and realized that I could figure out the gain setting such that, given the mic sensitivity and SPL handling specs, the input and the mic would simultaneously overload. Normally, I add a bit of gain to that setting (peak SPL is extremely unlikely at most shows I attend) and turn on the MixPre's excellent analog limiters. To my knowledge, the limiters have never actually kicked in.

Good gain staging.  When setup so that the mic and input reach overload simultaneously I wonder if it was the microphones or recorder that dominated the electrical noise floor of the recording chain.. not that you would ever actually encounter that in your recordings

I note that the EIN of all 32bit float recorders seems to be somewhere around -130dBu.  Their max input level varies a bit more.

[Gutbucket]

Recent 32-bit experience, and why I don't intend to go back to 24 bit:

Quick loudness check in the wavelab "Analysis" menu showed the left track was almost exactly 20 dB lower (the 59W mics do not have pads, and tests of the mics on a 24 bit Tascam recorder showed the mics were pretty closely matched; the F6 settings for both channels were identical).  [..snip..]  I suspect that fiddling with the connections during setup, with phantom on, caused a transient that "fooled" the F6 A/D to record the left channel lower.  Do the 32-bit A/D converters work in 20 dB jumps? Is this scenario possible or plausible?

That was my immediate suspicion.  The 20dB difference could represent a "step" of the auto ranging display I suppose, or my first thought upon reading your description of events was that 20dB represents the available headroom present before the impulse exceeded the dynamic limits of the analog input stage and was clipped at that point.

So I boosted the left track 20 dB, and ended up with a perfect stereo pair, no sign of any artifacts from system or background noise, which is exactly what 32-bit is supposed to do.

 
Okay, but a 24 bit file created from the same multi-ranging ADC should work the same way, because the mathematical range available in a 24bit fixed point file also exceeds the total dynamic range capability of the recorder.   In either case, that unplugging transient would exceed the recorder's real world headroom limit before exceeding the limits of either file format.

I don't see any difference in this scenario if you had recorded in 24 bit rather than 32.  I do see a difference if you weren't employing the recorder's multiple ADC feature.

Thanks for the discussion guys.  Not trying to be a contrarian here.  I just want to better understand this and have never heard a really convincing argument for 32bit fp recording on its own terms. In contrast to that I totally get the advantage of a multiple ADC design.  In most folks minds those two things are one and the same, but they needn't be. Perhaps its because that's how these recorders are marketed.  I kind of suspect its a marketing thing.  32bit is a bigger number and so is interpreted as being better.  Too hard for a competitor to explain why its really not, as long as the total dynamic range capability of recorder itself remains the true bottleneck.

[live2496]

24-bit recording will give you a lower noise floor. So the advantage of that is that you can record at a lower level and allow yourself some more headroom.

[Gutbucket]

^ Only in some cases..

24-bit recording can provide a lower noise floor, yet will be able do so only if the noise floor which is a result of the combination of everything else in the signal chain is lower than the least significant couple bits of the 16bit file space.  If that is the case, the maximum potential gain in headroom that can be achieved without incurring an increase in noise-floor will be the difference between those things. 

If the noise floor of all that other stuff is only a bit or two lower, that's all the additional headroom you get.  If the noise floor of all that other stuff is higher than the least significant bit of the 16bit file space, you can gain no additional headroom and just end up recording four additional bits of random noise.. that is, if you were to switch from 16 to 24 bits and leave the gains where they were.  If you reduce gain to increase headroom, you end up raising the noise floor of the recording by that amount, and produce a recording with more than 8 additional bits of noise at the bottom.

If the intent of the brief post above was to sum up the discussion (at least up to the point where we diverged to discussion of multi-ADC designs and 32bit fp), it's difficult to improve on the summary quoted below form an earlier post in the thread-

[..snip] 24-bit recording offers you no greater capacity to record higher numeric values (representing greater sound pressure levels) than 16-bit; ±1 is still the absolute outer limit that you don't want to crash into. But up to a point, if your signal path is generally sensible, 24-bit recording can let you reduce your record levels enough so that you don't run out of room at the top, and still not pay a price in increased audible noise. Not always--your mikes themselves and/or your outboard preamp, if you use one, can be overloaded before the signals get to the recorder; again, the ceiling isn't being raised but the floor is being lowered, one may reasonably hope.

The amount that lowering is never anything like what the additional 8 bits would imply (48 dB!), though, because no A/D converter or analog input circuitry has that much dynamic range. One well-known recorder that I measured some years ago gave only <2 dB improvement when going from 16 to 24 under the particular conditions I was measuring it. It wasn't a typical scenario, but it does illustrate the real dependency that exists on all the other factors in your actual scenario. If someone had used that recorder with those signal levels and gain settings except that they thought, "With 24 bits I can reduce my levels by (say) 12 dB to be absolutely safe from overload", their 24-bit recording would then have been ~10 dB noisier than the 16-bit one. Again that is based on an atypical scenario in terms of signal levels and gain settings, but not impossible by any means.

Thus you should always use the input to the recorder and the menu and level settings, if any, that are most appropriate for the signal levels you're dealing with, and always try to get your absolute peak sample values somewhere in the top, say, 6 dB or so of the recorder's range without hitting 0. It's just that with 24 bits as opposed to 16 (and with everything else optimal, please note) there should be no audible "noise penalty" if your highest peaks are only at, say, -8 dB rather than -2. So by setting somewhat conservative levels when you don't know what's going to happen (i.e. life), you're better prepared in the event of the unpredictable. Murphy's Law says, after all, that the unpredictable is the most predictable thing there is.

[Gutbucket]

The details about all of this were discussed early in the thread if you care to go back and read through it.

[Gutbucket]

David, I'm hoping you might comment on the question I posted earlier which remains unresolved:

What am I missing / misunderstanding?  Why did manufacturers feel the need add the 32bit fp format when no 32bit fp recorder is capable of capturing a dynamic range greater than what can already be handled by the 24bit recording format?

[Rairun]

If you have multiple ADCs and you're using a fixed point format, wouldn't you have to assign the clipping point of the least sensitive one to 0 dBFS, and then the more sensitive ones to lower ranges of the digital format? So a Sound Devices MixPre II (dynamic range of 142 dB) would have to commit to recording very quiet sounds (converted by the more sensitive ADC, providing less analogue noise) at the lower end of its 142 dB dynamic range, so that it if it recorded an explosion immediately after, we would never get digital clipping? The problem is that the resulting file, if the sounds recorded were quiet enough, would sometimes need to be boosted by something like 100 dB in post, as we wouldn't be allowed to set the gain!

It seems to me that the advantage of using a 32 bit floating point format is that 0 dBFS is not clipping, and the digital noise floor is much lower anyhow, so you don't ever have to record the quieter sounds anywhere near it.

[DSatz]

Gutbucket, you asked for my reply to this:

> What am I missing here?  Why did manufacturers feel the need add the 32bit fp format when no 32bit fp recorder is capable of capturing a dynamic range greater than what can be handled by a 24bit recording format?

I'd say that it's business economics (in a market segment different from the one that most of us are in) mixed with behavioral economics. It's straightforward business economics in the market for $20,000 recorders for film sound, where 32-bit recording offers advantages like those we discussed w/r/t digital microphones a month or so ago. It's also a good fit when the production sound crew hands its work off for post-production and mixing on computer workstations. The data conversion step that it saves is small but still greater than zero. It's no mystery why 32-bit recording is used in that work environment.

When and if 32-bit recording offers real advantages for the average person doing live music recording, however, it isn't because of 32-bitness as such, but because of everything else that WOULD go with it in an optimal implementation. It can just as well be implemented in ways that offer no advantage to anyone, as an empty selling point. If a $300 recorder really had similar capabilities to a $20,000 recorder, word of that fact would seep out sooner or later, I'd imagine. Back in the 1980s plenty of CDs were made from PCM-F1 recordings; despite certain problems, most of what was needed was there. I don't see any similar situation happening with low-cost 32-bit audio recorders, though.

"Behavioral economics" controls the world we actually live in. It's the study of the difference between (a) the way people hypothetically MIGHT make choices if they had all the relevant information about the value of something, could interpret it properly, and had no conflicting psychological baggage versus (b) the way people DO make choices, given fragmentary information of uncertain quality and a whole slew of biases that they're in denial about. Marketers and other business executives study this field closely; social media platforms make many billions of dollars from putting it to use; as I write this, some people on this board are very likely wearing clothes with outward-facing insignia and/or brand names that they paid to wear, rather than being paid to wear them. Supply and demand still exist, but corporations also create demand--often by letting customers feel that their individual identity and personality are being expressed by representing the brand.

The difference between (a) and (b) above is famously much bigger in consumer audio than in any other technical hobby or field. Few amateur astronomers would spend five times as much for telescopes based on the supposed properties of their exterior paint; therefore such telescopes aren't generally offered to them. Few ham radio operators would imagine that signals sent through a transceiver carrying the insignia of a long-dead brand will reach farther than one with identical measured performance but no such brand; therefore such transceivers aren't generally offered to them. In consumer audio, however, we're totally swamped with such things. Even pro audio has bent in that direction over the time I've been in it. It's profitable to cater to the credulous; significant opportunities are lost if a manufacturer fails to do so.

--best regards

[Gutbucket]

^
Thanks for that.  Good to know I'm not nuts.  At least with regard to thinking that through.

Whenever I eventually end up purchasing one of these multi-ADC design recorders I suppose I'll need to do some testing to make sure I could actually use that particular recorder to record in 24bit mode using the same recording modality as 32bit fp operation, thus avoiding the increased storage space requirements and the need for format conversion back to fixed point.

[Gutbucket]

The problem is that the resulting file, if the sounds recorded were quiet enough, would sometimes need to be boosted by something like 100 dB in post, as we wouldn't be allowed to set the gain!

That describes 32bit float recording!  Problem or feature?

It seems to me that the advantage of using a 32 bit floating point format is that 0 dBFS is not clipping, and the digital noise floor is much lower anyhow, so you don't ever have to record the quieter sounds anywhere near it.

 

Not an actual advantage because the bottleneck is not the 24 bit fixed-point file format, but the clipping and noise floor thresholds of the signal path through the recorder upstream of file format.  In this situation, both 24bit fixed and 32bit floating point formats are similarly capable.

[admkrk]

... and the need for format conversion back to fixed point.

If I am not mistaken, most DAWs work in 32 bit float, so there is no conversion in that stage and the conversion to a lower bit depth is going to happen either way when rendering. Peter's set I posted earlier today spiked at +3.? dB. Listening to it from Reaper, it did not seem to clip, but did over power the headphones. Lowering the gain bellow 0 dB got rid of the overpowering and still provided a full wave form.

One of the mics was 20 dB off, and I compensated for that, but it had other issues as well. It is interesting that someone else had the same offset regardless the cause for it.

[Rairun]

Not an actual advantage because the bottleneck is not the 24 bit fixed-point file format, but the clipping and noise floor thresholds of the signal path through the recorder upstream of file format.  In this situation, both 24bit fixed and 32bit floating point formats are similarly capable.

Is noise floor not cumulative, though? Meaning that when you add 100 dB to a recording, you're boosting both the noise floor of the signal path AND the noise floor inherent to the 24 bit fixed-point format?

I've just recorded 10 seconds of audio with my laptop's webcam mic, and silenced the audio on audacity. Once the audio is silenced, you can amplify it by however much you want, and the result is still going to be silence (internally, Audacity works with 32 bit float). Then I did the following experiment:

(1) I exported the audio to signed 16 bit and 24 bit PCM format.

(2) I opened the 16 bit file and boosted it by 50 dB. The noise level became very significant!

(3) I opened the 24 bit file and boosted it by 50 dB. It still sounded completely silent. I boosted it by 50 dB again, and then the noise became very significant!

Now, yes, if the noise floor of the signal chain is boosted by the same amount, you're right  that it will be louder than the noise I heard in (3). But surely the total noise level will be a combination of both noise floors, and not just the highest one? This is usually pretty irrelevant because we users do reasonable gain staging, so the signal chain's noise floor is always much, much louder than the 24 bit noise floor. But if your recorder always needs to record the quietest sounds at something silly like -140 dB (to make sure a sudden, deafening noise will not clip), then suddenly both noise floors will be pretty similar, and adding one to the other will result in a noticeably louder noise once boosted!