The Advantages of 32 bit floating point

[silhouette]

I remember a while ago being told that it was pointless recording at 32 Bit Floating point as it offered no sonic advantages. Maybe this video and article might be interesting?

A short video

The 32-Bit Float Workflow - Sound Devices

The 32-Bit Float Workflow - Sound Devices

A more in depth written explanation, but not that long.
32-Bit Float Files Explained - Sound Devices

32-Bit Float Files Explained - Sound Devices

It seems that the increase in dynamic range is truly worth it.
Needless to say I did not take any notice of the advice and continued to work in 32 bit floating point. It is interesting to learn that if you record very quietly, no matter how much you raise the clip level the noise floor does not increase.

[cubic13]

From what I understand, it’s essentially the ability to manage without digital distorsion clipped items that is interesting. IMO, raising the item level will also raise the noise floor, as essentially, the noise is related to the incoming signal : ambience noises, not optimal S/N ratio of devices used to capture the sound, the noise floor of the audio device itself being negligible, these days…

But even at this point, I’m wondering if using 32 bits floating point clips is really useful as, AFAICS, nearly all audio devices are still doing something like a 20 bits conversion of incoming signals (it’s my case, at least). So, I’m still using 24 bits, but I’m not a specialist, I admit…

[silhouette]

If I am not mistaken what I am taking from some of the literature that I have read the increased dynamic range means that when you raise the clip level the noise floor does not increase noticeably.
Quote:
The dynamic range that can be represented by a 32-bit (floating point) file is 1528 dB. Since the greatest difference in sound pressure on Earth can be about 210 dB, from anechoic chamber to massive shockwave, 1528 dB is far beyond what will ever be required to represent acoustical sound amplitude in a computer file.

There is one other aspect of 32-bit float files which is not immediately obvious. Files recorded with 32-bit float record sound where 0 dBFS of the 32-bit file lines up with 0 dBFS of the 24- or 16-bit file. Keep in mind that unlike the 24- or 16-bit files, the 32-bit file goes up to +770 dBFS. So compared to a 24-bit WAV file, the 32-bit float WAV file has 770 dB more headroom.

[cubic13]

All this is very well, and I’m trying to follow, but I obviously still have a problem with 32 bits FP files for recording.

I understand (correct me if I’m wrong) that a non clipping event will be recorded as something like :

24 bits		 : 011011010001100111100101
32 bits FP	 : 011011010001100111100101 00000000

And a slightly clipping one, as :

24 bits		 : 111111111111111111111111 -> distorsion, as the possible value recordable is maxed, no matter the real one.
32 bits FP	 : 000000000000010101110101 00000001

The latter is preserving the signal integrity. Overall, a real advantage of 32 bits FP. So, to avoid any clipping hassle, it seems a no brainer : let’s use it.
|

BUT…

A) What I don’t get is that I have never seen an audio device with ADC working in 32 bits FP. I just checked : even high-end converters such as Lynx Aurora, Prism Atlas or RME M-32 are all working at 24 bits. So, what happens when a too hot signal hits the DAW at the hardware AD conversion stage ? And how Cubase Record file format setting at 32 bits FP could fix anything, in this case ? This is, rightly or wrongly, one of the reason I continue to use 24 bits fixed point format, being careful, of course, to avoid any clipping (I usually much more have the other issue around , with mics not delivering enough levels…).

B) I have seen in Cubase 11 pro Operation manual the following (p.105), concerning the Record file format :

When you record with effects, consider setting the bit depth to 32 bit fl oat or
64 bit fl oat. This prevents clipping (digital distortion) in the recorded fi les and
keeps the audio quality very high. Effect processing and level or EQ changes in
the input channel are done in 32-bit fl oat or 64-bit fl oat format, depending on
the Processing Precision setting in the Studio Setup dialog. If you record at
16 bit or 24 bit, the audio will be converted to this lower bit depth when it is
written to a fi le. As a result, the signal may degrade.

If I am still understanding well, the minimal audio processing format in Cubase is 32 bits FP, so no clipping can internally occur through multiple audio processing operations. But this format has a 24 bits mantissa representing a given sample and an 8 bits exponent, the part designed to prevent any clipping. So, I guess that, when any rendering is done at 24 bits, the whole mantissa representing this sample value is put to file. From which, I don’t understand the part of the extract I put in bold : when any degradation is supposed to occur, when having a Record file format set at 24 bits (16 bits is another story, as there IS obviously a conversion, in this case) ?
|

It’s not to be confrontational. I simply still don’t get the need for 32 bits FP format files, beside the clipping prevention while doing multiple audio processing operations internally…
|

EDIT : OK, my reasoning is wrong in B), in the case where a clipping has occured. In this case, rendering only the mantissa wouldn’t give an accurate reproduction of a clipped signal, as the exponent value should also be taken into account.
So, I take back the whole B) paragraph. Still have a problem with A), though…

[silhouette]

Some good points here that I have thought about. Given that my soundcard is an RME UC with a 24 bit ADC. I am hoping that someone out there will be able to explain clearly where I am wrong in my thinking, or indeed where I am right. I have been using 32 bit FP for quite a while and have been very happy with the dynamic range and the handling of clips. However I do wonder if I have misunderstood the nature of the internal workings of Cubase with the input/recording of audio and the limitations of the soundcard.

[peakae]

Yes the point of using 32bit float file format is to be able to reduce the level after saving/rendering and get a signal that is not clipped.
Saving a audio file that has a signal above digital 0dB FS , and being able turn the signal down without it being destroyed.
I think of it as a cork in a glass tube, the cork has 24bit, now I fill the tube with water and it floats up. If it gets above 0dB FS then the converter clips. But I still have the full 24bit preserved, so I can turn down the level. I would say most users don’t bother and use 24 bit, and take care not to add so much level that the signal exceeds 0dB digital. It really just comes into play if doing stuff like Render In Place and doing it with to hot levels. Internally Cubase already uses 32bitfloat so it is only a problem when going to disk.
I myself use 32bit float since forever, simply because the file size difference between 24bit and 32bit float is not that big. And should I mess up, I would be able to correct my mistake. Have I ever needed it ?
I can not remember if I ever did,.
I am far more likely to record a signal to hot and have it clipped on the way in, and no file format is going to save a already damaged signal.

[cubic13]

@silhouette

…
Given that my soundcard is an RME UC with a 24 bit ADC. I am hoping that someone out there will be able to explain clearly where I am wrong in my thinking, or indeed where I am right. I have been using 32 bit FP for quite a while and have been very happy with the dynamic range and the handling of clips. However I do wonder if I have misunderstood the nature of the internal workings of Cubase with the input/recording of audio and the limitations of the soundcard.

I’m also using an RME device (Fireface UCX - USB connection). And yes, I think that the audio device limitation is the key, at least at the tracking stage. After this, maybe it is more relevent to use 32 bits FP when a lot of audio processing is needed, without bothering about gain staging. It would be interesting, though, to have a Steinberg crew member opinion on this, as the Operation manual isn’t exactly relevent, especially when no clipping occurs : how does Cubase renders 32 bits FP processing to 24 bits fixed point files, and is there any actual degradation in the process ?

@peakae

…
I am far more likely to record a signal to hot and have it clipped on the way in, and no file format is going to save a already damaged signal.

Yep… And we are returning again to the used audio device capacities. It’s a common and wise idea when recording is involved : better try to make things as clean as possible at the tracking stage than try to fix an already crippled record.

At the end, am I wrong when saying that, as long as we are careful with the levels used in the audio processing chain (“gain staging”), the 32 bits FP format for recording is useless ?

[silhouette]

This is an interesting article dealing with the same issue:

adventures-in-audio.com - This website is for...

[cubic13]

Indeed, and thank for the link. What I’m retaining from it is this :

But that doesn’t matter because a 32-bit floating-point file has no more resolution than a normal 24-bit file.

This shows us that unlike conventional fixed point 24-bit audio, the resolution and dynamic range are different things.

32-bit float achieves its amazing dynamic range by taking a 24-bit ‘window’ and scaling it up or down as needed. So if the level is high, the other eight bits scale the 24 bits to higher values. If the level is low it scales them down. But it has no better accuracy or precision than a 24-bit file recorded with a good gain setting.

I think that the excerpts put in bold more or less sum up the whole debate : all is about a relevent “gain staging”…

[silhouette]

I too would like someone from Steinberg to add a comment upon this as it would be of great interest to me. I have found that running in 32 Bit FP has had no downsides for me. However, I don’t know if it has had any upsides.

[MrSoundman]

The genius of VST was the realisation that the almost redundant “math coprocessor” (or “FPU”, floating-point unit, which used to be a separate chip) could be used for audio processing once the (then) 8- or 16-bit integer samples from an interface were stored as 32-bit floating-point values.

The whole “in the box” thing is born from the fact that consumer-level computers attained this wonderful capability (which almost none of them used at the time) when the FPU became integrated with the CPU as we now know it.

The limitation is where this interfaces with the analog world. I’m not aware of any A/D or D/A converters that use more that 24 integer bits, for the very good reason that nobody could possibly hear a difference; anything else is marketing*.

“In the box” – use whatever you need to in order to achieve your objectives, but remember that whatever format you ultimately render to, it needs to be readable in 20, 30, 50, 100 years … what you do in between, is your own business.

*lies

[Johnny_Moneto]

To keep it simple: Record audio with the maximum quality your ADC can deliver. If your audio interface has effects built-in look in its manual in which quality the driver passes data to Cubase and match that.
For bounces and mixdowns and any edits to audio where new files are created use at least 32 bit float. Always.
For the final mix get your audio levels in line with whatever is the your required output quality (often 16 bit)

This works in 100% of all cases and you can stip to worry about it.

[silhouette]

Thanks for that. Yes it is precisely that. I bounce down, or should I say Render In Place a lot and for that reason us 32 bits. Once I am happy with the midi instrument parts I render to audio, especially if they are short parts. I rarely use freeze these days.

[RJI-777]

IMHO, there are a couple of astute observations in Johnny’s post.

@Johnny_Moneto

To keep it simple: : Always keep it simple.

Record audio with the maximum quality your ADC can deliver. : I produced my first multitrack recording circa. 1974 using 2×cassette recorders, 2×mics, 2×c6o tapes, etc. The process was extremely limited, the sound quality poor and I don’t recall if the i/o was 24-, 16-, 8-, …, whatever bit [sarc]. However, when it comes to my digital recording, it’s all done at 24-bits.

If your audio interface has effects built-in look in its manual in which quality the driver passes data to Cubase and match that: : Sounds logical but, personally, very rarely do I record FX, other than the initial shaping of the instrument’s primary waveform e.g., guitar + distortion. In fact, bass is often recorded direct as it’s the sound of my bass I want; nothing else. As a rule, EQ and all FX are applied after a decent audio signal has been committed to disk.

For bounces and mixdowns and any edits to audio where new files are created use at least 32 bit float. Always… : I agree with the 32(+) bit float argument for internal processing e.g., FX, being higher than the i/o bits, although techniques such as “bouncing”, “audio editing”, constitute a completely different argument. Re final mixdown, see below.

For the final mix get your audio levels in line with whatever is the your required output quality (often 16 bit). : Again, in essence, I agree; but… how you configure your “final mix” settings to enable the provision of an optimal consumer audio product (and, yes, it is often 16-bit) is way beyond the scope of this thread. However, all I will say is that you need to understand the purpose of your final mixdown. This might appear simplistic, but is it? e.g., I could argue that the final mixdown levels should be in line with the requirements of the mastering process which, in turn, needs to consider what will happen to the digital master as it is reconstituted from, say, a high quality 24-bit WAV to… A what? Seriously?.. No way… A 16-bit MP3!.. How does that work exactly? Then again, does it matter? Or, in other words, do your individual listeners even care?

This works in 100% of all cases and you can stip to [sic] worry about it. : I think this relates to Johnny’s first point i.e., keep it simple, which I interpret as being consistent; again, key to working “in 1oo% of all cases”.

In summary, but at the risk of trivialising the process: Keep it simple and be consistent.

[cubic13]

@RJI-777

…
In summary, but at the risk of trivialising the procesd: Keep it simple and be consistent.

Exactly what I do : 24 bits for everything (included bounce, render, etc.) until final mixdown (usually 16 bits with dithering). Simply because, again, 32 bits FP format is actually 24 bits with an added part (exponent) that allows sloppy levels adjustments, at the expence of 34% more bigger files…

Still wondering why a dynamic range of 144 dB can’t be enough - I’ll keep myself from evoking the ol’ 24 (or less) tracks tape days…

[vncvr]

there’s been a couple times I was broad stroke gainstaging events in 6db increments and because I was zoomed out I simply didn’t see that some peaks had gone over simply because they didn’t graphically render in properly.

Because I was working in 32bit, all I had to do was turn down those parts back down and the peaks were in still intact, they didn’t clip.

Other instances, I wanted to roll back some of the stuff I had on the masterbuss including peak limiting and let the mastering engineer handle it… which meant for sure there would be some overs - didn’t matter. Just sent the mastering engineer the 32bit file.

also, the higher the dynamic range… the lower the noise floor no? A lot of plugins generate noise

[cubic13]

@vncvr

there’s been a couple times I was broad stroke gainstaging events in 6db increments and because I was zoomed out I simply didn’t see that some peaks had gone over simply because they didn’t graphically render in properly.

Because I was working in 32bit, all I had to do was turn down those parts back down and the peaks were in still intact, they didn’t clip.

Well… Usually, when something clip, it’s hearable, especially using digital audio stuff. So, I’m afraid that I don’t get your point. And I would be curious to know which plug-in/gear only allows level adjustements in only 6 dB increments.

If levels cannot be adjusted correctly, then fine : indeed, 32 bits FP is your friend. I’ve never been in such a situation, so…

@vncvr

Other instances, I wanted to roll back some of the stuff I had on the masterbuss including peak limiting and let the mastering engineer handle it… which meant for sure there would be some overs - didn’t matter. Just sent the mastering engineer the 32bit file.

In this case, OK, if he accepted to cope with it (guess it’s a part of his job, I don’t know). But AFAIK, usually, we avoid any clipping in a final mixdown before the mastering process, precisely to ease it.

@vncvr

also, the higher the dynamic range… the lower the noise floor no? A lot of plugins generate noise

144 dB… The ‘dynamic range’ of the human ear is, at its best (young ones at 1kHz) 120 dB and even a signal at 90 dB SPL will more or less quickly damage our ears when applied too longer : this is equivalent to a full orchastra playing. Here, the maximal exposition accepted by law in public is 80 dB SPL : guess where is the -144 dB noise floor of a 24 bits digital recording system, in this scale…

Beside this, and I’ll take an amp sim as a typical rather noisy plug-in : I suggest you to try to render a guitar track with an amp sim both at 24 bits and 32 bits FP. Ther will be no difference, because the noise generated by it is independent of the intrinsic noise generated by the digital system we are using. The mantissa of a given sample in the 32 bits FP file will have the same value as the corresponding sample in the 24 bits one, from which we will get exactly the same amount of noise from a recorded guitar + amp sim track in both cases.

[vncvr]

@cubic13

Well… Usually, when something clip, it’s hearable, especially using digital audio stuff. So, I’m afraid that I don’t get your point. And I would be curious to know which plug-in/gear only allows level adjustements in only 6 dB increments.

If levels cannot be adjusted correctly, then fine : indeed, 32 bits FP is your friend. I’ve never been in such a situation, so…

it’s not a plugin, it’s a DOP preset I’ve set to +6db and another -6db which is a perfect bit and thus has no resolution loss/rounding distortions.

Not all clips are audible. the point I was making is that I had destructive rendered an audio event in which the peak had exceeded what would be the typical ceiling and would be squared off, except it wasn’t because it was a 32bit fp file - I just had to pull the volume down and the peak was still intact. You wouldn’t be able to do that with 24bit.

@cubic13

In this case, OK, if he accepted to cope with it (guess it’s a part of his job, I don’t know). But AFAIK, usually, we avoid any clipping in a final mixdown before the mastering process, precisely to ease it.

the point is, there was no clipping because it was 32bit, and I pretty much resent him the same levels just without the peak control I had put on, so he was able to essentially keep his settings the same on his end with some minor tweaks.

even outside of this context, I can just generally work quickly ballparking levels to be fairly hot, sitting in the ballpark of where the master will be, and not worry about the peaks that are “just” clipping the master… because they won’t actually be clipping. ie, with 32bit FP, the clip light becomes more of a tool to gauge than it is something to worry about in terms of signal destruction. I can worry about those peaks later or let the mastering engineer handle it. there’s two paths here - turn all the material down and master things back up to volume… or… keep the volume the same and control the peaks.

@cubic13

Beside this, and I’ll take an amp sim as a typical rather noisy plug-in : I suggest you to try to render a guitar track with an amp sim both at 24 bits and 32 bits FP. Ther will be no difference, because the noise generated by it is independent of the intrinsic noise generated by the digital system we are using. The mantissa of a given sample in the 32 bits FP file will have the same value as the corresponding sample in the 24 bits one, from which we will get exactly the same amount of noise from a recorded guitar + amp sim track in both cases.

Low Level Signals: 32-bit Float versus 24-bit - Sound Devices

I believe also volume resolution/fader resolution is affected, which is important because as you turn something down you are losing resolution, more bits would theoretically offer greater rounding accuracy.

This isn’t about how loud it can go

[cubic13]

@vncvr

it’s not a plugin, it’s a DOP preset I’ve set to +6db and another -6db which is a perfect bit and thus has no resolution loss/rounding distortions.

Not all clips are audible. the point I was making is that I had destructive rendered an audio event in which the peak had exceeded what would be the typical ceiling and would be squared off, except it wasn’t because it was a 32bit fp file - I just had to pull the volume down and the peak was still intact. You wouldn’t be able to do that with 24bit.

Sorry, but… What is a DOP preset ?
Beside this, and I already stated it, sloppy levels adjustements are allowed with 32 bits FP and I’m glad that you’ve been able to recover what was at stake. I have no actual problem with this : we just use our respective DAW differently. For me, this doesn’t justify the 34% bigger files and the added strain on it.

@vncvr

the point is, there was no clipping because it was 32bit, and I pretty much resent him the same levels just without the peak control I had put on, so he was able to essentially keep his settings the same on his end with some minor tweaks.

even outside of this context, I can just generally work quickly ballparking levels to be fairly hot, sitting in the ballpark of where the master will be, and not worry about the peaks that are “just” clipping the master… because they won’t actually be clipping. ie, with 32bit FP, the clip light becomes more of a tool to gauge than it is something to worry about in terms of signal destruction. I can worry about those peaks later or let the mastering engineer handle it. there’s two paths here - turn all the material down and master things back up to volume… or… keep the volume the same and control the peaks.

Not sure that I got your reasoning : is it so difficult to adjust the export level, before delivering it to the mastering engineer ?

@vncvr

Low Level Signals: 32-bit Float versus 24-bit - Sound Devices

LOL ! First, we are not dealing with low level signals - as far as I have retained from 40 years experience and playing - it’s rather the contrary, when guitars are involved. Beside this, what we see in your link are beautiful graphs with signals having been boosted to… 100 dB !

How are these supposed to have with any actual user situation ? And honestly, did you just once have to boost a recorded signal up to 100 dB ? If so, there is truely a problem with your setup. So, the guy does and he sees that there is a difference : all is well and 32 bits FP files are the Graal of any audio rendering. Point taken : I’ll just let each of us consider in which way this can be transposed to a day-to-day practice…

@vncvr

I believe also volume resolution/fader resolution is affected, which is important because as you turn something down you are losing resolution, more bits would theoretically offer greater rounding accuracy.

This isn’t about how loud it can go

So, better to make everything clip as needed, if pushing your reasoning : 32 bits FP is in the rescue. Fine…
By the way, it isn’t about ‘more bits’ : see again how 32 bits FP format is structured : a 24 bits mantissa and an 8 bits exponent. So, it’s probably better to stop the exchange, at this point…

[vncvr]

@cubic13

Sorry, but… What is a DOP preset ?
Beside this, and I already stated it, sloppy levels adjustements are allowed with 32 bits FP and I’m glad that you’ve been able to recover what was at stake. I have no actual problem with this : we just use our respective DAW differently. For me, this doesn’t justify the 34% bigger files and the added strain on it.

Actually, I’m extremely tidy with my gainstaging, which is why I do +6 / -6db increments at the audio event level, sometimes there are some needle peaks that get in the way of the rest of the material being at an optimal level - I either graphically miss them sometimes, or, I just don’t care because I know it’s not an issue and a compressor will handle it later on, or I will craft the reduction by hand when the time to do so in the session is appropriate.

That’s not sloppy mate, that’s just being aware and utilizing the technology I have available.

DOP = direct offline processing.

@cubic13

Not sure that I got your reasoning : is it so difficult to adjust the export level, before delivering it to the mastering engineer ?

the mastering engineer already tweaked their gear, if I changed my levels to drastically, it would changing everything on their end. I took off my peak limiting, which resulted in faux-clipping, and they handled the peaks on their end.

@cubic13

LOL ! First, we are not dealing with low level signals - as far as I have retained from 40 years experience and playing - it’s rather the contrary, when guitars are involved. Beside this, what we see in your link are beautiful graphs with signals having been boosted to… 100 dB !

How are these supposed to have with any actual user situation ? And honestly, did you just once have to boost a recorded signal up to 100 dB ? If so, there is truely a problem with your setup. So, the guy does and he sees that there is a difference : all is well and 32 bits FP files are the Graal of any audio rendering. Point taken : I’ll just let each of us consider in which way this can be transposed to a day-to-day practice…

How do you know who is dealing with what??? I deal with low levels in situations all the time, sessions are full of happy accidents when things are leveled for drums maybe, and someone starts strumming something on a guitar during lunch. Sound design, creating convolutions, etc.

@cubic13

So, better to make everything clip as needed, if pushing your reasoning : 32 bits FP is in the rescue. So, it’s probably better to stop the exchange, at this point…

Someone sounds like they lost an argument