Topic: 32Bit Float recording - The Technical view

[Gutbucket]

Rairun- You've got it right.  There are two or more parallel analog circuit paths (call them preamps if you like) into separate ADCs, each analog path having a different fixed gain, and the output of those ADCs are combined into a single output.  That's essentially it.  In modern properly designed "32bit recorders" the combining happens in real time while making the recording.  But the same can be done afterward if the different ADC outputs are saved separately.  That's what TheJez's cool SafetyTrackMerger (Windows + MacOS) tool does for older "pre 32bit float era" recorders which have a feature that records a "safety track" at a lower gain simultaneously with the "primary track". They use essentially the same method of parallel analog circuit paths with different gains feeding multiple ADC channels. Biggest difference while recording is that you'd still manually set the gain for the primary channel and usually the offset gain on the safety track as well.  With modern 32bit recorders that gain setting is determined by the manufacturer in the design phase and locked down.  SafetyTrackMerger then does the combining part afterward on the computer.

TheJez is spot on when he says "multi ADC (if implemented properly) can and will broaden the analog dynamic input range capability".  No question about that. Its made possible by to the parallel circuit paths.

Voltronic is too when he confirms what I was saying about the dynamic bottle-neck being the analog input stage, where improvements in circuit performance are more costly.  The SD788 he mentions (now old!) and especially Grawk's Sonoxax are an examples farther along that axis of input stage quality.

If twice as many input stage paths or more running in parallel are needed that gets more costly.  But combining a number of less costly paths, each of which in isolation doesn't provide premium performance, may be able to achieve premium performance in aggregate.  And that's a reasonable way to increase performance while keeping costs manageable.  Use multiple premium input circuits and you can potentially increase performance dramatically in a costs-be-damned type of design.

But regardless of how the sausage is made, the proof is in the eating.  How this stuff works internally is interesting, but as I taper I don't really care as long as the device delivers the performance that is promised. The problem as I see it is that "promise".  The performance of these inexpensive 32bit recorders does not deliver the inferred (and sometimes outright) promises made in the marketing.  And measurements of what the actual real world performance is tends to be buried, withheld or obscured.

Long live independent testers with the capability to run the tests that can shed light on what's actually being delivered. 

/Data not promises.

[Rairun]

Right! I'm particularly interested in the architecture of the Zoom H1essential because it's suspected that they are using the same preamp as the Zoom H1 and H1n. I'm very familiar with the Zoom H1, which has a gain scale of 0 to 100, where <16 is digital attenuation, 16 is no gain, 17 to 36 is digital gain in steps of +0.4 dB, 37 is fixed (and very clean!) analogue gain of +13.5 dB, and >37 is digital gain in steps of +0.4 dB (on top of the +13.5 dB analogue gain). In other words: Zoom H1 and H1n users have a binary choice between +0 and +13.5 dB analogue gain. There is no reason to use anything other than levels 16 and 37.

At level 37, a very loud rock show would clip with the Zoom H1 internal mics, so you need to set it to 16. In my experience, a pair of CA-11s + battery box into a Zoom H1 would never clip at level 37, just because the CA-11s are less sensitive. So I can totally see the Zoom H1essential using less sensitive mics, fixing gain at +13.5 dB (which is clean, if that's all the gain you need) and writing this to a 32-bit float file. Or they could simply use the same mics and skip the preamp. When used with internal mics, either option would sound noisier than the Zoom H1 when recording quiet sounds, but it wouldn't clip with louder ones. When used with a less sensitive external mic, the former would sound as good as the Zoom H1, and the latter would sound noisier, with no risk of clipping either way. If used with a more sensitive external mic, it'd either clip more easily or have more noise during quiet passages.

Personally, if that's the case, I'd much prefer the Zoom H1 over the Zoom H1e. All the H1e would be doing is pretending you don't have to set levels anymore, by simply setting it to a non-ideal level for you!

This is assuming, of course, that the Zoom H1essential DOES NOT have two ADCs. It's very possible that it does, despite the lack of info in the specs. Assuming the same preamp as the H1/H1n and dual ADC, the device would give you the same performance as the Zoom H1 for quiet sounds (good if you only need +13.5 dB; if you need more, you're screwed), but with more headroom.

[Ozpeter]

OK, the game is up.  I have just received the following email from Tascam USA customer support -

"Hello Peter,
 
Our engineering department has sent the following information...
 
DR-07XP/DR-05XP are SINGLE ADC.
 
 
Thank you for your interest in TASCAM products."

So this confirms my suspicion that low cost 32 bit float recorders which do not claim dual A/D converters do not have them.  But they are happy to allow consumers to assume - like the Tascam marketing department person did when replying to me on YouTube- that dual converters are always used in 32 bit float devices.  These single A/D devices are probably just expressing their 24 bit output as 32 bit float, to gain advantages in the final result which are very hard to identify.

I have several 32 bit float devices, and the only one which does not claim dual converters is the H2essential.  It's fine for what it is, at the price they sell it for.  The lack of recording level control simply means that they have preset the level to cover the worst possible loudness scenario and the results are ok imho.  But given that all discussion of 32 bit float recording methods on the internet etc has described its benefits as being due to the use of dual converters - or more - it seems to be that companies now releasing single converter devices should include this fact in their published specifications which customers can see, and consider, before purchase.

[Rairun]

I have several 32 bit float devices, and the only one which does not claim dual converters is the H2essential.  It's fine for what it is, at the price they sell it for.  The lack of recording level control simply means that they have preset the level to cover the worst possible loudness scenario and the results are ok imho.  But given that all discussion of 32 bit float recording methods on the internet etc has described its benefits as being due to the use of dual converters - or more - it seems to be that companies now releasing single converter devices should include this fact in their published specifications which customers can see, and consider, before purchase.

Yeah. At that point, they aren't removing the need for gain control, they're just removing the OPTION to control gain in different circumstances, noise floor be damned.

[Ozpeter]

Even 'Deepseek' is deeply (!) confused on this subject - partly because it will be deriving its response from human stuff on the internet.

"You're absolutely correct that a single analog-to-digital converter (ADC) operating at 24-bit resolution cannot inherently capture more than 24 bits of dynamic range, and expressing its output in 32-bit float format doesn't magically increase the ADC's resolution. However, there are still practical benefits to using 32-bit float format even when the ADC itself is limited to 24 bits. Let me clarify the nuances:

---

1. **Preservation of Headroom and Avoiding Clipping**

   - While the ADC itself is limited to 24 bits, the 32-bit float format allows for **extended headroom** in the digital domain. This means that if the ADC's output exceeds 0 dBFS (full scale), the excess can be represented in the 32-bit float format without clipping.
   - In a fixed-point 24-bit system, any signal exceeding 0 dBFS would clip irreversibly. In 32-bit float, the extra "room" above 0 dBFS allows you to recover and adjust levels in post-production, even if the ADC's output was technically clipped.

[I think that is wrong in every respect]

---

2. **Simplified Gain Staging**

   - With 32-bit float, you don't need to worry as much about setting the input gain precisely. If the signal is too hot or too quiet, you can adjust it in post-production without degrading the audio quality.
   - In a 24-bit fixed-point system, you must carefully set the gain to avoid clipping or excessive noise, which can be challenging in dynamic recording environments.

[Wrong again, if the digitally clipped output from an overloaded 24 bit converter is stored in 32 bit float]

---

3. **Post-Processing Flexibility**

   - Even though the ADC's resolution is 24 bits, the 32-bit float format provides a much larger "container" for mathematical operations during post-processing. This reduces the risk of rounding errors or quantization noise when applying gain changes, EQ, compression, or other effects.
   - In a 24-bit fixed-point system, repeated processing can introduce cumulative errors, whereas 32-bit float minimizes these issues.

[OK, I'd agree with that]
---

4. **Future-Proofing and Workflow Consistency**

   - Using 32-bit float as the recording format ensures consistency with modern workflows, where 32-bit float is increasingly the standard for digital audio workstations (DAWs) and post-production.
   - Even though the ADC's resolution is 24 bits, recording in 32-bit float avoids the need to convert formats later, which can introduce unnecessary complexity or artifacts.

[That I'd broadly agree with too]

---
5. **Dynamic Range Utilization**

   - While the ADC's dynamic range is limited to 24 bits, the 32-bit float format ensures that the full range of the ADC's output is preserved. This is particularly useful for capturing very quiet signals, as the 32-bit float format can represent extremely small values without losing precision.
   - In a 24-bit fixed-point system, very low-level signals might be lost in the noise floor, but 32-bit float ensures that these signals are retained for potential recovery in post-production.

[Wrong if the 24 bit output from a single 24 bit converter is being stored in 32 bit float]
---

6. **Practical Example**

   - Imagine recording a live concert where the dynamics are unpredictable. With a 24-bit fixed-point system, you might set the gain conservatively to avoid clipping, but this could result in quiet sections being too close to the noise floor.
   - With 32-bit float, you can set the gain more aggressively, knowing that any clipping can be recovered in post-production, and quiet sections will still be preserved with full precision.

[Simply not true without multiple converters making more use of the 32 bit float than a single converter]

---

### Conclusion
While it's true that a 24-bit ADC cannot capture more than 24 bits of dynamic range, expressing its output in 32-bit float format provides significant practical advantages in terms of headroom [wrong], post-processing flexibility [maybe], and workflow efficiency [true if post-processing in 32 bit float, but not really a big deal]. The 32-bit float format acts as a "safety net" for the ADC's output, ensuring that the full potential of the 24-bit resolution is utilized without the risk of clipping or loss of detail [Not true]. This makes it a valuable choice for professional audio recording, even when using a single ADC.  [I'd say that a single converter system writing 32 bit float output might make the designers think carefully about end-to-end gain staging but that's probably the only good thing]

[GroundHog420]

There a point at which my brain apparently isn't translating all this information very well, but I reckon that's no one's problem but mine.

This may or may not relate to what some of you folks have been explaining, but on the user end, it's been frustrating over the past few years, that some audio applications automatically default to 32-bit, and there doesn’t seem to be any way to discern what the original format was.

Since some people have been encoding at variable formats such as 24/44.1, 24.48, 24/96, I can’t be sure any more when I get files what they were in the first place, and half the time, the people who sent me the files don’t really have any idea either, since they either got them from someone else, or were just following instructions they didn’t really understand.

I'm attaching a few screen snaps of files which I wouldn't know how to decipher if I hadn't been provided with the information first.

[morst]

4. **Future-Proofing and Workflow Consistency**

   - Using 32-bit float as the recording format ensures consistency with modern workflows, where 32-bit float is increasingly the standard for digital audio workstations (DAWs) and post-production.
   - Even though the ADC's resolution is 24 bits, recording in 32-bit float avoids the need to convert formats later, which can introduce unnecessary complexity or artifacts.

[That I'd broadly agree with too]

Apple Logic Pro has been 64 bit internal since 2013...

[Ozpeter]

Groundhog420, try this, which I have used for years without my PC being hijacked -

https://mediaarea.net/en/MediaInfo

Or maybe you are using an Apple device - well, it may be useful to someone anyway.

[Gutbucket]

4. **Future-Proofing and Workflow Consistency**

   - Using 32-bit float as the recording format ensures consistency with modern workflows, where 32-bit float is increasingly the standard for digital audio workstations (DAWs) and post-production.
   - Even though the ADC's resolution is 24 bits, recording in 32-bit float avoids the need to convert formats later, which can introduce unnecessary complexity or artifacts.

[That I'd broadly agree with too]

I don't agree with that.

I don't buy the workflow argument.  I can import whatever file format into the DAW without any extra steps.
No way fixed point WAV format is going away despite adoption of 32-bit float.  It will remain supported.
Converting from 24 bit to 32bit float is not complex at all and introduces NO artifacts. Just plain wrong on that account.


Ugh with AI's regurgitating erroneous human output.  One of the biggest problems with AI as I see it is an underlying misjudgement of presumed authority - the assumption by users that what it spits out is actually true!  I see that leading to dangerous and misleading simplifications in society at large as it becomes more widely adopted.

[Melanie]

Ugh with AI's regurgitating erroneous human output.  One of the biggest problems with AI as I see it is an underlying misjudgement of presumed authority - the assumption by users that what it spits out is actually true!  I see that leading to dangerous and misleading simplifications in society at large as it becomes more widely adopted.

I agree whole heartedly. Bob

[morst]

Ugh with AI's regurgitating erroneous human output.  One of the biggest problems with AI as I see it is an underlying misjudgement of presumed authority - the assumption by users that what it spits out is actually true!  I see that leading to dangerous and misleading simplifications in society at large as it becomes more widely adopted.

I agree whole heartedly. Bob

100%.


And there's a rule of thumb which lays it out...
https://thebullshitmachines.com/lesson-2-the-nature-of-bullshit/index.html

As we will see, LLMs are powerful tools. But they also make it easy for people to mislead us by accident, or on purpose.
The Bullshit Asymmetry Principle — also known as Brandolini's Law after the computer programmer who proposed it — is one of the most important principles in bullshit studies.
Brandolini's Law: The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it.

I think we have an intuitive understanding that this is the case... It's why the Galt Gallop works so well (just keep lying and they'll never catch up to you)

[Gutbucket]

 

Thanks for that link!

[Ozpeter]

Well - ok, I take your point(s).  But the nub of my post was to indicate the level of confusion on this subject - even in the world of AI.

Meanwhile Tascam have posted a remarkable response to my questions on YouTube about whether their new recorders, which use 32 bit float but do not claim dual converters, do or do not have single or dual converters.  Their "Product Specialist", who claims to know more than anyone in the company, states categorically that a single ADC cannot write its output in 32 bit float format, and therefore the devices must use dual converters.  He (or she) also doubts my statement that their engineering department states equally categorically that the devices have single converters.  The tone of the response is borderline offensive.  I have pasted into my response the text of the email I received quoting their engineering department and suggested that someone further up the company should explain the company's employees diametrically opposite statements and apparent lack of product and technical knowledge.

The confusion in the company and the manner of their communications makes me very reluctant to buy one of their products again.  Not that I planned to.  I will post the entire exchange in the thread here related to these new models.

[morst]

The confusion in the company and the manner of their communications makes me very reluctant to buy one of their products again.  Not that I planned to.

🗑
#TrashCan

[grawk]

Tascam and zoom both have similar strategies. They make some consumer focused gear which is mediocre and some pro focused gear which is significantly better. If you can afford to step up do it. If not, you can still make good recordings, it just requires more of you as the recordist.