32 bit float

GroupDIY Audio Forum

Help Support GroupDIY Audio Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.
Rather misleading / inaccurate imo.  When recording your converter will put a limit on max and min levels, 24 bit vs 32 bit float storage makes no difference, you can't even achieve true 24 bit performance at this time

32 bit float can make a difference in post processing, but even there the dynamic range claims are overblown.  It's a sliding scale,  you can store really large and really small numbers.  But you can't simultaneously manipulate really large and really small numbers.
 
On a side note, I would make the argument that fixed point (double precision or better) is actually superior for audio. We are dealing with a known relatively fixed dynamic range, and fixed point also allows proper dithering which is critical for high quality audio.
 
john12ax7 said:
Rather misleading / inaccurate imo.  When recording your converter will put a limit on max and min levels, 24 bit vs 32 bit float storage makes no difference, you can't even achieve true 24 bit performance at this time

32 bit float can make a difference in post processing, but even there the dynamic range claims are overblown.  It's a sliding scale,  you can store really large and really small numbers.  But you can't simultaneously manipulate really large and really small numbers.

funny fact! considering that their whole sale strategy is based on it  ;D
 
john12ax7 said:
Rather misleading / inaccurate imo. 

What do you find misleading?

Look at a bit meter in 32 float.  You'll see the advantage in both processing and storage if you think about it.  Unless all you do is record loud pop with no dynamic range, then you won't see any advantage.  Yes, no sense capturing most converters at anything more than 24, BUT there are now 32f converters out there on portable recorders that capture mic signals at unity gain as well as those using similar dual ADC's for higher dynamic ranges.  Set ideal gain after the fact.  Give it a couple of years, and it'll be in everything. 
 
That first paragraph seems dicey. It reads:

"There is in fact so much headroom that from a fidelity standpoint, it doesn’t matter where gains are set while recording. Audio levels in the 32-bit float WAV file can be adjusted up or down after recording with most major DAW software with no added noise or distortion."

If gain is low you can amplify it digitally later and, yes, technically you are not adding noise. But the existing noise will be amplified! If gain is higher, the noise floor in final proper level-adjusted recording could end up being lower if the noise floor if the analog amp is better than that of the circuitry leading into the ADC.

And of course the format of samples doesn't really matter that much once you get to 24 bits. My guess would be that the only real advantage to using floats or 32 bit samples is so that you don't have to convert if your hardware or software is already using that internally.
 
Their claim only makes sense with the other claim that their mic pres are so low in noise, the additional noise (from boosting digitally after the fact) will not matter.
I have the mix pre II but I´m only using it in 24 bit for now- The 32 bit is useful but only for very specialised tasks(wide/unknown bandwith) or for the totally clueless ;)
 
EmRR said:
What do you find misleading?

Thought I explained it? A converters circuitry would determine its dynamic range.  You don't magically gain headroom by choosing to represent the converter data as float vs fixed.
 
john12ax7 said:
Rather misleading / inaccurate imo.  When recording your converter will put a limit on max and min levels, 24 bit vs 32 bit float storage makes no difference, you can't even achieve true 24 bit performance at this time

32 bit float can make a difference in post processing, but even there the dynamic range claims are overblown.  It's a sliding scale,  you can store really large and really small numbers.  But you can't simultaneously manipulate really large and really small numbers.
You can manipulate really large and really small floating point values at the same time but there is little point since the small stuff gets swamped by the big stuff...

The extra resolution for small stuff is useful for handling small coefficients in digital LF filters running at high clock rates, involving very tiny amounts added/subtracted each clock tick...

We mere mortals can't perceive that much dynamic range, but it sounds good for FAB sheets (features, advantages, benefits).

JR
 
john12ax7 said:
Thought I explained it? A converters circuitry would determine its dynamic range.  You don't magically gain headroom by choosing to represent the converter data as float vs fixed.

Well, there are multiple types of headroom.  This is but one.  If you can’t anticipate the dynamic range you might encounter, this offers more latitude. 
 
EmRR said:
Well, there are multiple types of headroom.  This is but one.  If you can’t anticipate the dynamic range you might encounter, this offers more latitude.

Could you explain what you are referring to? What are the multiple types of headroom for the front end capture?

(I think we can all agree more bits are useful for DSP)
 
john12ax7 said:
Could you explain what you are referring to? What are the multiple types of headroom for the front end capture?

(I think we can all agree more bits are useful for DSP)

OK, yeah, headroom is really not the right word. 

If you effectively erase any S/N barrier from the digital portion, you have more options within the analog domain, and within digital signal processing of those signals.  Capture raw mic outputs at low levels, effectively adds headroom because the nominal signal level has been vastly lowered without fear of being close to where dither might live. 

I've had clients bring me sessions in which there was no recorded level at all, no gain added with preamps, virtually no meter movement.  Turn it up enough to mix and hear, listen to the dither and hash at the bottom.  Sure, not a common mistake, and I've never understood how they monitored it in the first place. 

32f in theory allows you to capture the crickets in the field and the jet flyover with no gain adjustment at capture.  Ramp the levels in post.  Yeah, not many people need that, some do, all may eventually benefit.  Really calls for a different analog front end design I would guess. 

It would seem forensic cleanup of noise floor artifacts might produce better results with the digital noise floor further out of the way. 

A larger format camera. 

etc.

early days. 

 
Actually it seems to me that there's one major caveat here.

If the mic pre and converters are one device (like Sound Devices recorders), then yes, once the mic pre noise floor exceeds the converter noise floor, there is no noise performance benefit. In fact, it is potentially a disadvantage because the signal could be clipped or distorted. You can amplify digitally later with no impact on noise performance.

But! If you have any analog gear in between (and that could just be the output stage of a pre in a separate box) the noise floor of that is going to be relatively high which necessitates that you boost the signal enough to get good noise performance from it.
 
squarewave said:
Actually it seems to me that there's one major caveat here.

If the mic pre and converters are one device (like Sound Devices recorders), then yes, once the mic pre noise floor exceeds the converter noise floor, there is no noise performance benefit. In fact, it is potentially a disadvantage because the signal could be clipped or distorted. You can amplify digitally later with no impact on noise performance.
It is not trivial to maintain an input (analog) noise floor lower than the A/D convertors noise floor "and" take advantage of the digital dynamic range (upper bits). Premium convertors will dither the LSB trading the grainy quantization noise for more natural sounding noise floor (but still noise). Digitally boosting (multiplying) a signal encoded too low will amplify the digital noise floor. This sounds better than the digital noise floor of old technology convertors but while the gain does not add any noise of its own, there will always be a noise floor.

But! If you have any analog gear in between (and that could just be the output stage of a pre in a separate box) the noise floor of that is going to be relatively high which necessitates that you boost the signal enough to get good noise performance from it.

Good (old) rule of thumb is that gain should be applied as early in the signal chain as possible, so later gain doesn't boost the noise of the earlier stages.

JR

PS: About the only free lunch in digital processing is that you can combine an unlimited number of channels without signal degradation. Of course if you exceed the dynamic range of your medium you need to scale down the output signal, but this further reduces the noise floor for even better S/N. Floating point digital data allows us to represent dynamic range far exceeding our needs or ability to use as final output, but can be beneficial for number crunching while inside the digital domain.
 
Lets assume that you have  24 bit converter with a maximum input (analog level) of +22dBu, thats 9.76VRMS or a peak voltage of 13.798V, That means that the lowest step (the LSB) the converter can code is 13.798/(2^24) = 0.82 microvolts, thats -122.5dBU (after converting to RMS) of minimimum signal, thats the noise of a 1Kohm resistor across a 20KHz bandwidth, giving a roughly 144.5 dB of dynamic range, most microphones or line level devices are no way near that noise floor, which basically means that the first 3 or 4 bits in your 24 bit converter are always on, effectively making your 24bit converter into a 20 bit or so converter. So whats the point of having a 32bit converter? I do agree that 32 bit float in post processing provides an advantage, but in the conversion stage?  I dont think so, to me 20bit is more than enough.
 
user 37518 said:
Lets assume that you have  24 bit converter with a maximum input (analog level) of +22dBu, thats 9.76VRMS or a peak voltage of 13.798V, That means that the lowest step (the LSB) the converter can code is 13.798/(2^24) = 0.82 microvolts, thats -122.5dBU (after converting to RMS) of minimimum signal, thats the noise of a 1Kohm resistor across a 20KHz bandwidth, giving a roughly 144.5 dB of dynamic range, most microphones or line level devices are no way near that noise floor, which basically means that the first 3 or 4 bits in your 24 bit converter are always on, effectively making your 24bit converter into a 20 bit or so converter. So whats the point of having a 32bit converter? I do agree that 32 bit float in post processing provides an advantage, but in the conversion stage?  I dont think so, to me 20bit is more than enough.
More is always better right?  ::)

In theory a small enough sine wave would get quantized into a square wave but in practice the noise dithers the sine wave providing apparent resolution down below the noise floor.

IMO modern digital conversion has become pretty much transparent.

JR
 
Sound Devices has patented 32 bit processing...

When I asked one of their representatives how it was different from previous designs, he pointed me to the patent and a few of their marketing pages. These are not clear on the subject. When I asked him, again, what he thought was different about their approach, he ran with his tail between his legs.

Is this a pure marketing patent or will we see Zoom Japan et al paying licenses to SD?
 
Sound Devices are not the only ones with a patent here, they are circumventing someone else's patent as I understand it.  Then Zoom are doing yet another thing. 
 

Latest posts

Back
Top