32 Bit Float Explained

[antcollinet]

I'm a trifle confused here. I understand that the sampled data is being converted to and stored as a 32-bit float. However, isn't the analog-to-digital converter only going to be working at 24-bit integer precision?

Possibly - or it might be more.

Either way, doing any processing in 32 bit float eliminates of any rounding errors and overflow errors. After DSP it will be converted back to integer of whatever resolution the DAC part of the system requires.

 

[alexanderino]

What many detractors of the 32-bit float format miss is the added benefit of being able to boost levels without raising the noise floor. It isn't just about recovering 'clipped' audio (which also is a thing of the past with 32-bit float, as far as the format is concerned). 16-bit and 24-bit fixed point offer 96.3 dB and 144.5 dB ranges respectively. 32-bit float, by contrast, contains a whopping 1528 dB of dynamic range. If any clipping/distortion occurs, it is due to hitting the mic's (or pre-amp's or other components') limit, poor micing technique, etc. — not the format itself.

See for yourself how superior it is in practical terms; view the following clip from 5:38 to 9:29:

There are two immediate benefits:

1. (5:38–7:30) Signals that cross the 0 dBFS threshold (but are within the mic's range) can be recovered without any loss. 24-bit or 16-bit fixed point cannot recover signals above 0 dBFS and remove the clipping introduced by the format's ceiling itself.
2. (7:59–9:29) Signals that are recorded too low can be amplified to a usable level without pulling up the noise floor as well (which is -758 dB in 32-bit float, with another 770 dB above the nominal 0 dB point). In the video, Judd boosts an extremely weak section of the signal by a whopping +98 dB and still ends up with a perfectly usably low noise floor. This is simply not achievable with 24-bit or 16-bit fixed point.

The cost for such versatility? An extra 33% storage space, which may sound considerable but is a drop in the ocean in this day and age of multi-terabyte drives. With 32-bit float, I can ignore recording level knobs and limiters entirely if I so choose. Of course, if I so choose, I can continue to practice good rec level setting techniques that are critical with fixed-point formats but not 32-bit float. No more recordings ruined due to incorrect gain settings, unexpected loud/quiet sounds in uncontrolled environments, etc.

The one criticism that can be legitimate is the larger file sizes. If the added 33% is an issue, however, one can always convert the final product to 24-bit or 16-bit to recover space while ensuring no clipping or noise floor hiss is present.

32-bit float is a game-changer, because the recording format is no longer a relevant limiting factor in practical terms. I for one am quite excited by, and love working with, it.

 

[AnalogSteph]

Of course, the digital guys wanted to know why we couldn't just design a 32-bit, 24 GS/s ADC and grab X-band signals directly. In ~1985. Pretty sure I couldn't do it even today.

LOL. How ludicrous. If I am not mistaken, the resistive noise level of a 12 GHz bandwidth at room temperature by itself would be at -93ish dBm. I just checked what you can actually buy these days, and found some 10 and 10.2 GSPS 12-bit parts (you don't even want to know the SFDR)... just 3 grand a pop, what a bargain. And people are complaining about the prices of high-end audio ADCs...

 

[DonH56]

LOL. How ludicrous. If I am not mistaken, the resistive noise level of a 12 GHz bandwidth at room temperature by itself would be at -93ish dBm. I just checked what you can actually buy these days, and found some 10 and 10.2 GSPS 12-bit parts (you don't even want to know the SFDR)... just 3 grand a pop, what a bargain. And people are complaining about the prices of high-end audio ADCs...

Yah. I was working with systems having 140 to 160 dB dynamic range with 10 to 100 MHz bandwidth after processing (converters running at around 1 to 10 GS/s, but those were 6-10 bit converters back then, about 40 years ago). The problem we had, and an issue with floating-point converters, was we were trying to deal with very weak signals in the presence of very strong signals so FP was not a lot of help except in the processing. Performance was limited by the data converters, albeit we were doing some tricks to improve their resolution (colored dither, oversampling, analog and digital filtering, analog gain ranging where we could, I/Q converters to increase system bandwidth, etc.)

 

[bennetng]

What many detractors of the 32-bit float format miss is the added benefit of being able to boost levels without raising the noise floor. It isn't just about recovering 'clipped' audio (which also is a thing of the past with 32-bit float, as far as the format is concerned). 16-bit and 24-bit fixed point offer 96.3 dB and 144.5 dB ranges respectively. 32-bit float, by contrast, contains a whopping 1528 dB of dynamic range. If any clipping/distortion occurs, it is due to hitting the mic's (or pre-amp's or other components') limit, poor micing technique, etc. — not the format itself.

See for yourself how superior it is in practical terms; view the following clip from 5:38 to 9:29:

There are two immediate benefits:

1. (5:38–7:30) Signals that cross the 0 dBFS threshold (but are within the mic's range) can be recovered without any loss. 24-bit or 16-bit fixed point cannot recover signals above 0 dBFS and remove the clipping introduced by the format's ceiling itself.
2. (7:59–9:29) Signals that are recorded too low can be amplified to a usable level without pulling up the noise floor as well (which is -758 dB in 32-bit float, with another 770 dB above the nominal 0 dB point). In the video, Judd boosts an extremely weak section of the signal by a whopping +98 dB and still ends up with a perfectly usably low noise floor. This is simply not achievable with 24-bit or 16-bit fixed point.

The cost for such versatility? An extra 33% storage space, which may sound considerable but is a drop in the ocean in this day and age of multi-terabyte drives. With 32-bit float, I can ignore recording level knobs and limiters entirely if I so choose. Of course, if I so choose, I can continue to practice good rec level setting techniques that are critical with fixed-point formats but not 32-bit float. No more recordings ruined due to incorrect gain settings, unexpected loud/quiet sounds in uncontrolled environments, etc.

The one criticism that can be legitimate is the larger file sizes. If the added 33% is an issue, however, one can always convert the final product to 24-bit or 16-bit to recover space while ensuring no clipping or noise floor hiss is present.

32-bit float is a game-changer, because the recording format is no longer a relevant limiting factor in practical terms. I for one am quite excited by, and love working with, it.

While floating point processing is the typical standard of DAW when dealing with data in digital domain, remember everything before digital conversion, including mic and preamp are still limited by law of physics and your 1528dB will become useless. Just read the previous posts in this thread.

 

[BeerBear]

What many detractors of the 32-bit float format miss is the added benefit of being able to boost levels without raising the noise floor...

If only, that would be a miracle... When the guy in the video says that it was recorded at -60dB, ask yourself what that really means.

If the Zoom F6 is like that other Zoom device we mentioned earlier, it doesn't have user adjustable gain. Instead, the "input level" affects an already digitized signal. So the signal was captured at a fixed gain, converted to digital and only after that reduced by 60dB by the user. That's why you can boost it back by 60dB without increasing the noise floor. But anything more than that increases the noise floor that exists in the recording. And that noise floor is also limited by the device's dynamic range, which as we've seen earlier might be around 140dB (which is already great, really).


So, again, I think the advantage of these 32bit float recordings is just that it allows setting a more user friendly level, because the manufacturer can already set the 0dBFS point (in the device) to a level that's more relevant for normal usage. Whereas if they used an integer format, 0dBFS would have to be at the very top level of gain to prevent clipping and the regular level recordings would therefore appear low in volume. But the actual quality and noise floors of the recordings would be the same (if using sufficient bit depth, of course).

 

[danadam]

Nice, we were promoted from confused to detractors. Or is it demotion?

Re. floating point itself, here are some online tools you can fiddle with bits inside floating point and see what happens:

IEEE-754 Floating Point Converter

Float Toy

 

[alexanderino]

While floating point processing is the typical standard of DAW when dealing with data in digital domain, remember everything before digital conversion, including mic and preamp are still limited by law of physics and your 1528dB will become useless. Just read the previous posts in this thread.

Which is why I noted:

32-bit float is a game-changer, because the recording format is no longer a relevant limiting factor in practical terms.

We won't need any more than 210 dB dynamic range to capture all possible sounds in nature (in fact, 210 dB will likely be lethal as the accompanying shockwave could kill us). However, does it hurt to have 1528 dB at our disposal? No. Is it detrimental to the final captured output? No.

Do you see the value in being able to rescue excessively loud/quiet sounds as seen the Curtis Judd video?

 

[earlevel]

I'm a trifle confused here. I understand that the sampled data is being converted to and stored as a 32-bit float. However, isn't the analog-to-digital converter only going to be working at 24-bit integer precision?

What some people might be missing is that this is a field recorder. not a studio recorder, not a live music recorder. With live or studio, you set level, for field recording, you might not have a chance to set levels. A field recorder might be used for recording birds sounds in the wilderness, sounds on a crowded street, etc.

Example: Say you're recording a protest for a video news team. You want to capture crowd chants from further away, with input gain (mic sensitivity) set accordingly. When it's time to interview someone, you adjust to their voice level. A truck drives by with an external speaker, blasting slogans. With a simple recorder, the latter is clipped. Consider, also, recording the distant sounds of birds in a quiet forest, and having an screeching eagle fly by—it would be pretty sad to have it clipped.

The traditional way around it is to have an automatic gain control pumping up and down at throughout your recording, which hurts fidelity—you've lost the true dynamics, and might have a pumping noise floor. But if you have multiple ADCs, you can simultaneously capture the same mic input at multiple settings without changes in gain, and you can still capture their relative loudness.

So, yes, you've captured at a wide dynamic range. But correct, you can't play it back with that dynamic range, unedited.

 

[bennetng]

Which is why I noted:


We won't need any more than 210 dB dynamic range to capture all possible sounds in nature (in fact, 210 dB will likely be lethal as the accompanying shockwave could kill us). However, does it hurt to have 1528 dB at our disposal? No. Is it detrimental to the final captured output? No.

Do you see the value in being able to rescue excessively loud/quiet sounds as seen the Curtis Judd video?

The sound pressure at this level already caused damage to the mic, so did your demo actually had 210dB SPL? Or even 140dB for this matter?
It doesn't matter of course, and DAWs are moving to 64-bit float anyway, but always mentioning whopping 1528dB sounds quite funny.

 

[bennetng]

I hoped that would be 8-bit floats, which they weren't if I read that correctly. Oh well, let's be generous and assume they meant something else, so that at least one part of their sentence is true.

(Just to be clear to anyone reading, my facepalm was for they implying that 8-bits sounds flat and robotic)

Never mind. My reply was not targeted at your comment, just to give some general insight for those who are unfamiliar but interested in this topic.

OPS on professional synthesizers was the premium version using more bits while other chips (e.g. OPL, OPN) were often used in vintage video games. However FM (and PSG) synths are not aimed at realistic sound regardless of bit-depth anyway. Something like this:

Many old games in 80s and early 90s only used very little amount of sampled audio because game data were often stored in expensive rom chips. Think about those Street Fighter game characters shouting those unintelligible "Hadouken" and "Sonic Boom" move names.

One of the earliest dedicated game console capable of offering CDDA quality audio was the NEC PC-Engine CD-ROM system in the late 80s. The discs are playable on traditional CD players and data were often stored on the second track, and the first track was usually a warning message that the second track contains non-audio data and can emit loud noise when being played, and the rest of the tracks can be played without issue. The biggest drawback is very slow loading time and fragile reading mechanism which is not really suitable for arcade use.

SNES in the early 90s also used rom cartridges and the largest games are only 48 Megabits (6 Megabytes).

Super Nintendo Entertainment System Game Pak - Wikipedia

en.wikipedia.org

Here is one of the 48 Megabits games reproducing fairly long voice dialog, which is very uncommon for this game console. It is an emulator recording instead of real hardware.

For comparison, rom size of Super Mario World on SNES is only 4 Megabits, so it is even possible to store 2 games on a single floppy disk, which was a popular way of pirating in that era. The audio processor of SNES (Sony SPC700) was mainly used for reproducing background music using sample-based synthesis. The attached file is an actual SNES analog output recording using the Sony MZ-R3 MD recorder, so not lossless but the overall noise level can still be identified.

 

[voodooless]

Never mind. My reply was not targeted at your comment, just to give some general insight for those who are unfamiliar but interested in this topic.

OPS on professional synthesizers was the premium version using more bits while other chips (e.g. OPL, OPN) were often used in vintage video games. However FM (and PSG) synths are not aimed at realistic sound regardless of bit-depth anyway. Something like this:

Many old games in 80s and early 90s only used very little amount of sampled audio because game data were often stored in expensive rom chips. Think about those Street Fighter game characters shouting those unintelligible "Hadouken" and "Sonic Boom" move names.

One of the earliest dedicated game console capable of offering CDDA quality audio was the NEC PC-Engine CD-ROM system in the late 80s. The discs are playable on traditional CD players and data were often stored on the second track, and the first track was usually a warning message that the second track contains non-audio data and can emit loud noise when being played, and the rest of the tracks can be played without issue. The biggest drawback is very slow loading time and fragile reading mechanism which is not really suitable for arcade use.

SNES in the early 90s also used rom cartridges and the largest games are only 48 Megabits (6 Megabytes).

Super Nintendo Entertainment System Game Pak - Wikipedia

en.wikipedia.org

Here is one of the 48 Megabits games reproducing fairly long voice dialog, which is very uncommon for this game console. It is an emulator recording instead of real hardware.

For comparison, rom size of Super Mario World on SNES is only 4 Megabits, so it is even possible to store 2 games on a single floppy disk, which was a popular way of pirating in that era. The audio processor of SNES (Sony SPC700) was mainly used for reproducing background music using sample-based synthesis. The attached file is an actual SNES analog output recording using the Sony MZ-R3 MD recorder, so not lossless but the overall noise level can still be identified.

I remember when I had a C64 with a disk that took an age to load, and when done, would play a sampled version of Wham’s Wake Me Up Before You Go-Go. I have no idea how they did that on the SID chip, but it was pretty cool back then

 

[BeerBear]

However, does it hurt to have 1528 dB at our disposal?

But we don't have 1528dB at our disposal, do we? When recording we're still limited by the dynamic range of the device. For the MixPre II series (32bit float) that's 142dBA. Manufacturer's words, not mine.

 

[MRC01]

OPS on professional synthesizers was the premium version using more bits while other chips (e.g. OPL, OPN) were often used in vintage video games. However FM (and PSG) synths are not aimed at realistic sound regardless of bit-depth anyway. Something like this:
... (insert ZeroWing video game clip) ...
Many old games in 80s and early 90s only used very little amount of sampled audio because game data were often stored in expensive rom chips.

Not to mention saving money with terrible translations to English, like the famous, "all your base are belong to us"!

All your base are belong to us - Wikipedia

en.wikipedia.org

 

[Multicore]

A very silly format from the perspective of an audio engineer.

Doesn't that depend on what is being engineered? For example Boss RC505ii uses it. Considering the difficulties of level setting for a device like that the dynamic range of 32 bit FP makes engineering sense to me.

 

[bennetng]

But we don't have 1528dB at our disposal, do we? When recording we're still limited by the dynamic range of the device. For the MixPre II series (32bit float) that's 142dBA. Manufacturer's words, not mine.

In fact I don't really mind the increased storage, what I don't like is that all the internal adjustments (switching between different ADCs) were performed on the fly and users don't really have any control about the processing after everything is mixed into a single file.

If people really care so much about these things then they should actually store two 24-bit copies in the recorder, and the recorder's manufacturer should provide a tool to process the recorded data in offline mode using a separate computer, and let users know where and how the processing was done, i.e. provide a preview using an automated preset and allows users to tweak the transition parameters and so on. Should also offer options to export to two separate audio files so that users can process the files with other third party tools.

See how deformed the recorded waveform after some uncontrollable processing happened:

Zoom F6 (32-bit float equipped)

Established in 2002, Taperssection.com is the forum community of live concert recordists. Taperssection is full of archived information and professional recordists who are ready to help you with your live sound recording questions

taperssection.com

Even low bitrate MP3 files don't suffer form these kinds of things.

Zoom F6 Portable Field Recorder Review

No, I did not find it out myself, it is advertised as such, for example in this review: https://www.solidstatesound.co.uk/sony_pcm-d100.htm As I said previously, I posted the full review archive with notes to describe the test procedure in this post...

www.audiosciencereview.com

 

[sarumbear]

Is Zoom the only manufacturer whose devices tested using 32-bit float? Zoom is the cheapest on the market, what about the industry standard Sound Devices? How about Tentacle’s micro wonder unit?

 

[BeerBear]

Right. Zoom messed up there, but I wouldn't condemn all multi-ADC designs because of that.
I'm sure there are challenges when stitching multiple converter outputs together, but maybe it's possible to do it right. And if so, great! Who doesn't like more dynamic range.
I don't really mind them using 32bit float either. The problem I have is with the deceptive marketing attached to it.

 

[sarumbear]

Right. Zoom messed up there, but I wouldn't condemn all multi-ADC designs because of that.
I'm sure there are challenges when stitching multiple converter outputs together, but maybe it's possible to do it right. And if so, great! Who doesn't like more dynamic range.
I don't really mind them using 32bit float either. The problem I have is with the deceptive marketing attached to it.

I don’t think there were any deception. It’s the standard not well engineered case that we see on ASR. They delivered what they promised but not fully.

 

[Cbdb2]

Really? A format that gives you the ability to record audio with no possibility of clipping without having to adjust the gain is silly?

In my view what is silly is commenting on a subject that you obviously have no grasp of it.

Here is a range of recorders that are created purely because of the format's existence.

32-Bit Float: Everything You Need to Know

zoomcorp.com

Worked recoding to 16bit devices for 20 years and never had a problem with clipping (or noise floor). Gain staging is recording 101. An so is limiting the dynamic before you record.
So basically recorders for people who don't know how to record.
And when is background noise not the limiting factor in dynamic range when recording a mic, even with 16bits.