32bit floating point recording question

Hi there ,

there is a lot of new recorders out there using 32bit floating point already for recording, like the zoom f6 or also the sound devices mix pre series ii

my questiog here is: how does these files behave in ardour?
ardour is surely capable of handling 32bit floating point (working internally like that anyways, right? )

but how do files behave recorded louder thatn 0dbFS , do they clip when played back too loud? would it be safe to normalize 32bit floating point files while importing to avoid that clipping? etcetc… maybe there are obvious answers to these questions but I m in the beginning of trying to understand all this…

2 Likes

I think the confusing part in the article is that does not stress the fact that an analog recording can not have a signal > 0dBFS.

This can only happen when summing two or more floating point values digitally.


Ardour can also record to 32bit float directly and all signal flow in Ardour is 32bit float.

You may have noticed that the default signal-meters in Ardour are up to +6dBFS. Signals inside Ardour cannot clip. This does not include plugins! Some effects may not support signals over 0dBFS, or other intentionally limit them.

Signals > 0dBFS are always clipped when they reach the boundary to the analog world. e.g. on playback to a soundcard.

“FS” = full-scale. 0dBFS means maxium signal the ADC/DAC supports.

Also when exporting to a fixed-point format. For the latter Ardour offers normalization (it’s on by default).

It’s safe (no information is lost), but there’s no benefit to doing that.

In music production one usually sums many tracks. So ideally the signal level of each individual track is low (say peak -15dBFS) and the peak of the summed result is already in the ballpark around -2dBFS.

The main advantage of recording to float is that you can use for lower levels. It provides ample headroom.

Some soundcards include specs at which levels ADC works best. A general rule of thumb: aim for around -20…-15dBFS when tracking.

1 Like

thanks so much for clarifying!!

my “import” question came right from that article, naively thinking that when there is something recorded to (exagerated case ) digitally 500dbfs I d kill my ears and speakers , though the signal is not clipping digitally :wink:

Just to be sure that I’m not misunderstanding, and, given that there is a MixPre-6 II winging its way to me as I type this, the value of the Zoom F6 and new MixPres is that not only is there 32-bit float format but also multiple AD convertors (2 in the case of Zoom, 3 in the SoundDevices) that essentially enable you to not even bother setting gain as different convertors contribute to different bit values of the 32-bit float file (or so the theory goes). Essentially, you only clip if your microphone clips. Am I accurate with this? There seems to be some debate about whether setting gain on the SoundDevices makes a subtle difference with hitting ADs at the optimum levels (and only detectable via a null test). The Zoom doesn’t give you that option as it records post-fader and thereby only varies the exponent, I think. Anyway, I’m excited by these new recorders as it means that as a last resort I could confidently show up last-minute and record a concert without worrying about where the gain is set.

Interesting. Stagetec has been doing something similar for many years, but it hasn’t really caught on because of the complexity and expense (Stagetec is using 4 converters per channel for the gain ranging, so essentially from a parts cost standpoint you could have 4 times as many channels for the same cost if you went with a traditional design).

Yes, I think that is the idea. There are always ways to poorly implement a good idea, but SoundDevices has a reputation for very high quality products, so I would expect it will work well.

I’ve heard something similar about the technology being around in the professional audio realm for a while. I think what is good about these particular devices is that I hope it pushes all popular makers of USB interfaces to try and follow suit (provided the Zoom F6 and MixPre ii are runaway successes) and give the option of ultra-wide gain. Of course. there would need to be a disclaimer to make sure 32-bit float format was being recorded to…

I suppose even if I show up early for a concert recording there would be no reason not to use the 32-bit float format but I’m still trying to understand whether setting gain will make any difference on the SoundDevices unit (do the knobs simply act as “exponent” faders in this mode?). Perhaps being able to change gain would defeat the purpose of having the three AD convertors in the first place set at optimum “crossover” positions? And, I would love to hear a normalized comparison between the 32-bit float (with gain/fader knob turned all the way down) and a 24-bit recording at healthy levels.

Why would there be any difference? 32 bit float is 24 bits with 8 bit exponent. So the same information is available either way. The float is great while mixing because over “zero” doesn’t clip. But exporting returns to 24 bits. Even with signals below the magic zero there is still 24 bits of info and so raising to almost zero for export still has 24 active bits. At the end of the day we are mixing for 9 or 10 bits (less in a car) and we package at 16bits… which is still more dynamic range than most home systems can handle and certainly well within the a studio background noise to painful range as well.

1 Like

Keep in mind that the thermal noise-floor at room-temperature is at about -127dBm (@48kHz bandwidth). You get at best 18-19 bits valid data from any analog sound-source.

The main benefit of 24bit audio or 32bit float is convenience, not fidelity.

2 Likes

@ Len
I would imagine there will be (inaudible maybe…) a difference depending on the design of the preamps. I still didnt get exactly how gainstaging happens with the mix pre II at 32 bit floating point. maybe the sound signature of the preamp will be more obvious with the levels cranked up? could be interesting to experiment…

The preamp has nothing to do with the number of bits, one is analog and the other is digital. In side the device is a 24bit ADC those 24 bits map directly to the 24bit part of the 32 bit float. That is a 24 bit int has exactly the same information as a 32 bit float so far as the static representation of an audio signal. The 32 bits float only becomes useful when processing the signal with effects or when mixing. The actual bit depth does not change but remains 24 bits of actual audio information. Most audio interfaces from a semipro on up have 24bit (int) ADC and that is mapped to a 32 bit float for use inside the computer. Ardour uses 32 bit floats throughout for all audio, even audio from 16 ADCs is converted to 32bit float. There really is no difference between 32 bit float and 24 bit int, they are a different representation of the same thing. That is, 24 bit int converted to 32 bit float converted to 24 bit int should be bit perfect. The use of 32bit float in this device allows keeping all 24 bits of information intact at different levels. That is 24 bits in from two or three ADCs at different levels can only be expressed with 32bit float while keeping 24 bits of information. Most ADCs are single and so can express the input directly with 24 bit int. In both cases the resolution is still 24 bits.

BTW imagining there will be a difference is not the same as there actually being one. The only way to find out is testing… from the math, I would expect that any measurable difference between this device and another would be in the analog components rather than the digital representation. This is not a put down of this device using 32bit floats. Using more than one ADC at different levels requires 32 bit float and that is what it would end up at in the computer anyway. I am just saying that it is still 24 bits of information. It is not a higher resolution or a lower noise floor (as Robin pointed out, the analog circuitry already limits the noise floor to less than 20bits of information) but rather an easier format for manipulating audio information.

I agree that 32-bit float is just 24-bit scaled up and down. The question here, I think, is whether the levels at which the mic input hits the preamps makes any difference to the sound quality. Again, it may not be noticeable but in others areas in the audio world it doesn’t stop people wanting the very best quality – think SRC and dither! Chris Johnson (airwindows) not only has what might be the world’s best dither (NJAD24 and 16 as part of StudioTan) but also a float dither both separately and baked into every one of his plugins. Could people tell in a blind test if 32-bit float dither is being used? Probably not, but it’s nice to know that using them is not negatively affecting the sound. And there’s the math to back it up. Same for SRC…Ardour’s Secret Rabbit Code seems great but it doesn’t stop programs like Brick, Smarc etc taking things to an extreme.

And, yes, testing is the only way. I would love to hear a recording of something acoustic, say a choral or symphonic concert, both in 24-bit and 32-bit float turned all the way down. As long as I don’t hear any difference over my monitors or hi-fi speakers I’ll gladly show up and use 32-bit float without even turning up the gain. If one sounds more sterile, then I’ll be sure to at least turn the knobs to unity gain. I’m not interested in null tests per se, but each to his own.

@ LEN yes I understand that of course this is all in the digital realm. but i can apply before different mic gains? if i can, then this will alter the sound, as mic preamps do usually not work 100%linear with every gain? and cant i saturate a mic preamp on purpose to get some sound colouring there? something i couldnt do to the extremes with 24 bit? there i was aiming at… but correct me when i m wrong, this is pretty new to me…

A device that has 32bit float outputs has a chain that looks like this:
Audio in->preamp with gain control->24bit int ADC->internal cpu->24bit->32bit float->to user’s computer. The gain control must be in the analog part of the circuit in order to make the maximum use of the bit depth without clipping. With two (or more) ADCs it would be possible to have the level control set two levels one you see and one that is hidden and exactly 20db down (or some other know value). The ADC at the higher level would be used for the output most of the time but if that adc goes over full scale the second one at 20 db down could be inserted by increasing it’s level by 20 db and switching to use that. The 32bit output would then faithfully represent the portion of the waveform that was over 0dbfs. However during that clip the bit depth would actually drop from 24 bits to 20-ish bits. It would still sound better than a digital clip. Also, it would allow recording at a higher level overall meaning a better use of the 24 bits available. However, it is questionable if regaining the last 4 bits is worth anything if the limits of the audio circuitry is only 19 to 20 bits anyway. There is not any way that using 32 bit float will help reveal preamp sound. The normal way to do that is to use a separate preamp that has the desired sound, set it’s level for that sound and then set the output level for correct ADC level in.

Nothing is “scaled” up or down, 24bit int and 32 bit float are the same information in a different format where each format will give the exact same output. Now it is true that once the data is reformatted to 32bit float it can be scaled up beyond limits imposed by 24bit int and that scaling down will not loose bits of information as the 24bit int does. But after conversion from one to the other there is no difference besides format until someone decides to modify the data. That is why most audio interfaces are raw 24bit int and most daws use 32bit float.

Of course it does… but that is not really the question at all. The reality is, in any recording case the preamp will be set in a similar position so that the peaks are less than 0dbfs and most of the audio is no more than about 20db down. Changing preamp level for colour is not in the cards in this case. Preamp colour means using some sort of extra level control after the preamp to ensure the signal level falls within the best recording range of the ADC. So the levels going into the ADC are pretty much constant for any circumstance. Dither is for exporting at 16 bit, not for signal processing a 32 bit float over and over as it goes through each effect in a chain. Dither is by definition noise that is added to the signal. If done right and only once it is useful, if done wrong and over and over the noise floor goes up and you loose bits of information. SRC is generally frowned on, it is only used where no other method will work (like those cheap USB microphones so many people want to use). SRC means something has failed and SRC is there to rescue an otherwise unusable situation.

What you hear over your monitors level wise, is not anything digital, but rather the analog level on the amplifier connected to the speakers.

This is what I meant!

I’m referring to the effect of having up to 3 different input levels (padded or otherwise) coming together to make up the 32-bit float file. If there are 3 different levels to accommodate a gain range extending beyond that of a good microphone, at least one of them will be hitting the preamp/convertor at a low level. I guess it doesn’t make too much of a difference in the real world though…

I disagree. If I record an orchestral concert at 96kHz/24-bit I always want to create FLAC and MP3s from that highest quality file and then down-sample (externally if necessary) to 44.1k in order to create a physical CD from the same markers. That’s standard practice and Wavelab includes a great custom montage feature just for that reason! Most people go to external SRC because the ones included in DAWs are generally worse but Ardour (Secret Rabbit), Wavelab (SoX) , Pyramix (Hepta Apodizing) etc are superb. There’s a reason people use Weiss Saracon on a regular basis – precisely to be able to master hi-res digital files at 96k or higher and then again at 44.1k for CDs using all the markers and edits they have already completed.

@ len thanks for clarifying, this all makes more sense to me now. so I d say for my usecases: recording music wont make a difference as i can set the gain as i wanted. when it is for set recording in a chaotic situation and a lot of channels 32bit floating point can make a huge difference.

https://people.xiph.org/~xiphmont/demo/neil-young.html
https://xiph.org/video/vid2.shtml
96k equates to twice the space, twice the noise, twice the cpu, distorts on most systems that can actually reproduce anything above 20khz.
Anyway, yes SRC on export may be useful for 48k to 44k1 or back (48k is standard and 44k1 is CD) But we were talking about audio interfaces, in the case of audio interfaces, SRC is a fix for a broken situation.

1 Like

Eek. You should probably tell this to Linn Records, a superb hi-fidelity classical label with perhaps the finest recording engineer alive to today, Phil Hobbs. They record at 192k for SACD and, I assume, down sample afterwards for the various formats including, on occasion, 44.1/16-bit harpsichord discs. SRC is far from a “fix for a broken situation”. They produce their own music systems and speakers too. I believe that in this conversation I personally have only mentioned SRC with regards down-sampling in software after capture.

Also, there are plenty of people who say that 96k or higher is essential for capturing complex reverb tails (and plus I find it is much better to SRC from a higher samplerate down to 44.1). I tend to be sceptical of ultra hi-res download options and often default to 44.1/16 for audio-only productions but a couple of people have convinced me recently that it is worth capturing at a higher sample rate at least (given HD space is cheap as chips these days). And, honestly, in conversations with Phil Hobbs while I was working in Glasgow, I trusted everything he said about audio and watching him coax a good performance out of an ill-prepared performer was magic in of itself. I digress…

The link len posted is my favorite when it comes to the sample rate discussion. Although i have to say there are moments 96kHz or beyond might make sense: some classical music recordists say it adds a silky “coloration” to the sound , that is favorable (so lets say high sample rate as a “pleasing effect”) [I have never witnessed that with my own ears.] or when you do fieldrecordings (bats) and want to analyse ultrasonics in baudline , or when you want to apply weird downward pitch shifts from crash cymbals for sound design. , etc…

for “general” production I don´t see any sense to it.

There is marketing and then there is physics. Marketing is about getting the highest dollar for a product. Physics is reality. Linn Records is selling a product for maximum profit. They are selling to the same people who are willing to pay extra for high sample rate reproduction equipment and gold speaker wire. They may feel they have to work in the fashion they do so that they can compete with other companies advertising high resolution audio. This does not mean anybody, and I do mean anybody, can tell the difference a higher sample rate than 48k makes. Very few people who have made past their teens can hear sounds as high as 18k, let alone 20k. Any difference heard is the separate mix done for the higher sample rate product… as in you’ve been hood winked.
Feel free to record and sell audio in any format you wish but do understand there is no audible difference aside from the greater distortion higher sample rate audio produces on high frequency playback equipment.