Zoom F8N

[twotoedsloth]

Hello,

Question for you:

Is the floating point mode on the Zoom F8N "clip immune"? I know that some of the other Zoom products have this feature, for example the UAC-232.

If it does have this capability, why are there still gain knobs?

While I would never tempt fate by consistently over-recording, it might be a lifesaver if one was recording, for example a percussion concert which has huge gain ranges - from whisper quiet to crash cymbals.

I haven't pulled the trigger yet, please let me know if I should!

Many thanks,

Peter

[Sam Spoons]

I'm guessing you will have read Hugh's review update but if not here it is. https://www.soundonsound.com/news/zoom- ... n-recorder

[Hugh Robjohns]

Is the floating point mode on the Zoom F8N "clip immune"?

Neither the original F8, nor the successor F8n offer floating-point recording. It's only the latest and current F8n-Pro model that does.

Just pointing that out in case you spied an attractive deal on an old-stock F8n...

But no, it's not 'clip immune' — no floating-point system ever is — because there is always some (high) signal level that will overload the analogue input stage.

However, the idea is that the front end should have sufficient headroom to cope with most real-world situations — achieved in part by applying minimal gain to the mic signal — while also having a very low noise floor courtesy of staggered A-D converters (one for loud bits, and one for quiet bits) which are combined in DSP to create the floating point file.

In essence, when using the 32bit floating point mode you are recording with a huge digital headroom which makes digital clipping highly unlikely, while quiet stuff is still captured with a good signal-noise ratio thanks to the staggered converters. And the actual captured audio level can be manipulated in the DAW afterwards to suit your production requirements.

If it does have this capability, why are there still gain knobs?

Partly because it also offers a 24bit fixed-point recording mode which requires manual gain optimisation, and partly because the knobs are used for mixing to a stereo track and for monitoring.

While I would never tempt fate by consistently over-recording, it might be a lifesaver if one was recording, for example a percussion concert which has huge gain ranges - from whisper quiet to crash cymbals.

One would hope you'd have a rehearsal to optimise gain settings for a concert recording... but yes, the 32bit floating-point mode is very helpful if you are recording in an unpredictable situation or if you simply aren't able to manually adjust the recorder for some reason.

I haven't pulled the trigger yet, please let me know if I should!

I can't say if you should buy the F8n-Pro. What I can say is that I like mine. It works very well, sounds pretty good, and represents excellent value for money. Technically, there are better quality recorders out there, but only at substantially higher prices, and they're only very slightly better...

What I would say, though, is don't forget to budget for some form of external rechargeable battery system, especially if you're planning mobile operation.

[Arpangel]

What "is" Floating Point? and what benefits does it bring?

[jjlonbass]

Floating point in computing terms is a means of increasing the range of numbers that can be expressed at the expense of precision.
In essence, floating point quantities have so much dynamic range that it is very difficult to achieve overflow or "clipping" when performing maths operations with them.

It helps to think about scientific notation for numbers as you will see on a scientific calculator. For example, 1E+03 as shown on a calculator is 1 multiplied by 10 to the power of 3 i.e. 1000. The part to the left of the E is known as the mantissa or significand and the part to the right is the exponent.
Floating point numbers in a computer are represented as a mantissa and exponent but both are expressed in binary and occupy a given number of bits of storage.

For example, a signed 32 bit integer can represent numbers between -2147483648 and +2147483647. A signed 32 bit floating point value with a 24 bit mantissa and 8 bit exponent can represent a maximum value of 3.4028235E+38.

John

[Arpangel]

Floating point in computing terms is a means of increasing the range of numbers that can be expressed at the expense of resolution.
In essence, floating point quantities have so much dynamic range that it is very difficult to achieve overflow or "clipping" when performing maths operations with them.

It helps to think about scientific notation for numbers as you will see on a scientific calculator. For example, 1E+03 as shown on a calculator is 1 multiplied by 10 to the power of 3 i.e. 1000. The part to the left of the E is known as the mantissa or significand and the part to the right is the exponent.
Floating point numbers in a computer are represented as a mantissa and exponent but both are expressed in binary and occupy a given number of bits of storage.

For example, a signed 32 bit integer can represent numbers between -2147483648 and +2147483647. A signed 32 bit floating point value with a 24 bit mantissa and 8 bit exponent can represent a maximum value of 3.4028235E+38.

John

Thank you for that! although I'm not going to say I understand any of it!
I think I need "The Observer Book Of Digital Audio"

[Drew Stephenson]

The simple, but probably wrong, explanation is that it's like a 24bit range with an extra 8 bits of extension that can be either higher or lower than the core range.

[RichardT]

Floating point refers to the position of the decimal point. Roughly speaking, floating point numbers have two components, one holding a number, the second indicating where the decimal point is to go.

This means that it can represent an extremely large range of numbers.

This is just a simplified, and not quite 100% accurate, version of Jjlonbass’ answer above.

[The Elf]

'at the expense of precision'... How so?

[BWC]

https://www.soundonsound.com/sound-advi ... nt-systems

https://www.soundonsound.com/sound-advi ... -explained

[jjlonbass]

'at the expense of precision'... How so?

Because the numerical value of the least significant bit of the mantissa i.e. its meaning if you like changes depending on the value of the exponent.

For example, if the exponent was 0, the least significant bit of the mantissa would have a numerical value of 1 just as if it were an integer.
If the exponent was 1, i.e. we're multiplying by 2 to the power of 1 = 2, the least significant bit of the mantissa would have a numerical value of 2. This means that we can not represent the value 1 or any odd number with this scheme, only 0, 2, 4 etc.
In practice, floating point numbers are "normalised" to make best use of the range available, but the change in precision is real.

John

[Wonks]

'at the expense of precision'... How so?

It's a mixed bag.

If you are talking about representing very large integers that fall within the 25 to 32 bit range, then you are loosing some precision at the lower end of the number with 32-bit floating point compared to a basic 32 bit number (anything in the bottom 8 bits of these large numbers - not numbers that can be represented by up to 24 bits) is lost, so it's slightly less accurate (but good enough for most purposes as it's a very small percentage of the overall number).

Note that a standard 24 bit number will normally use the first bit as the sign bit (showing if it's a positive or negative value) with the following 23 bits used for the number. However this is software specific and you don't have to use a sign bit if your software always deals with positive numbers.

Obviously any value larger than a 24-bit integer value can represent can't be represented properly, so you need say 32-bit floating point to represent a these larger values, albeit at the expense of some accuracy in the lower part of the number.

And whilst you can represent negative numbers using a 'sign' bit at the front of the number, you can't represent any decimals with a standard 8, 16 or 24 bit number without either:
a) using a set meaning within your software e.g. first 16 bits of a 24 bit value are whole numbers, last 8 bits are the decimal part - but that is very specific to the software and the values won't travel to anything else that doesn't use that convention or
b) you use some more numbers to show where the decimal point goes, which is where with 32 bit floating point, you get the first bit as the sign bit (saying if it's a +ve or -ve number), an 8 bit 'mantissa' showing where the decimal point goes and a 23 bit number for the value.

Of course you need to know if it's 32-bit floating or 32 bit fixed point when exporting the value to other software, but that's where the documentation comes in and 32 bit floating is a recognised standard.

64-bit floating is another standard and used by a lot of DAWs now, where you get the first bit as the sign bit, an 11 bit 'mantissa' showing where the decimal point goes and a 53 bit number for the value. So you can represent much larger values with more precision than with 32-bit floating point.

It takes more computing power to deal with 64-bit floating point values than 32-Bit floating point values, which is why 64-bit floating point wasn't used by DAWs to start with, but with today's powerful processors, it's not an issue.

[Hugh Robjohns]

'at the expense of precision'... How so?

Essentially because 123,456,789 is more precise than 1.234 x 10^8 — crudely comparing a 32bit fixed point representation with a 32bit floating-point (which has a limited wordlength for the audio data). Both numbers are equally large but the first has more precision in the tens and units columns which the second doesn't even register!

As others have said, in digital audio most converters code the signal with 24 bits in a 'fixed point' format where the audio amplitude indicated by each quantisation code is precisely defined.

This arrangement theoretically gives a dynamic range of 141dB (allowing for 3dB of dither noise) which is enough to code anything we're able to hear and tolerate in real life. The very best current A-D converters are actually delivering around 124dB dynamic range... so not quite living up to theoretical perfection... but still way more than most people need!

However... in post-production we are often mixing signals together and turning them down or up... and working with 24 bit numbers in a 24 bit accumulator is often going to result in overloads, or signals lost in noise — as can happen in a cheap another sound mixer.

The solution, as in the analogue world, is headroom. So one solution, adopted in early Yamaha digital mixers, was to use 32 bit processing inside the mixer, giving an internal dynamic range of about 192dB — more than the dynamic range of the atmosphere between double pressure and a vacuum (ie, the sound of an explosion!)

Other early digital systems used "double- or treble- precision" processing using 24 or 56 bits for the number crunching.

But the world of computing liked the idea of floating-point maths and so that was widely adopted by DAWs and is now the standard technique using 32bit numbers.

The way it works for digital audio is to allocate 24 bits for the audio, and 8 bits for the scaling. In that way, the potential dynamic range for signal processing is about 1500dB — a range so large you're never going to clip it or lose signals in the noisefloor in normal use.

In fixed point processing, if you turn a signal down it occupies fewer bits and so the signal-noise ratio worsens — just as in analogue. But with floating-point, when you turn the signal down you're only changing the scalar value — the original 24bit audio is retained intact... up until the point where you export the mix as a fixed point 24bit file.

Of course, 32bit floating-point isn't the only game in town. It was convenient when we had 32bit operating systems, and it has become a common audio file format when exchanging data between devices... but current operating systems are most 64bit and so 64bit floating-point maths is now standard in most DAWs. This allocates a longer wordlengths to both the audio chunk and the scaling section, so it gives even more dynamic range for processing and — importantly — even more precision in the audio itself.

Hope that helps...

[The Elf]

I'm clearly going to have to take much of this on trust. When I think of a floating point number I think of oodles of 0.0000000001 accuracy, so this all runs counter-intuitively to me.

[twotoedsloth]

Good Morning gentlemen,

Thanks for all of your replies, it has been fascinating.

I guess the main question between 32 bit floating point and 24 bit fixed, is, which one sounds better? I am guessing the 32 bit floating point, but if it is lacking precision, perhaps not...

Best regards!

[Wonks]

It only lacks precision in some certain specific circumstances. In others, withing the typical working range of audio samples, it has a lot more.

Remember that it is only when dealing with the output of the ADC that the 24 bits (or 16) are then converted to 32-bit floating point values. So until you start manipulating the sound files, the 32-bit float file has exactly the same value as a 'fixed' 24-bit file direct from the ADC would have.

And most DAWs convert to 32-bit float (and some have a 64-bit float options) to do internal calcs. It's only when you export a file you get back to fixed 24- or 16-bit files, which you need to put through your D/A converter. You can't do much about that (I've had a quick look on the web and there doesn't seem to be a DAC that accepts 32-bit float inputs - yet!).

[The Elf]

It only lacks precision in some certain specific circumstances. In others, within the typical working range of audio samples, it has a lot more.

[Hugh Robjohns]

I guess the main question between 32 bit floating point and 24 bit fixed, is, which one sounds better?

Ah... well... it all depends on what you're doing!

For example, if you're recording something very unpredictable, or simply can't get to optimise the gain structure, then the 32bit floating-point will definitely sound better than a 24bit fixed point recording which is clipped or lost in the noise floor.

However, there are reviews on the web that illustrate how the DSP which combines the outputs of the high and low converters can get caught out if a hugely loud transient sound occurs while you're recording something very quiet. The result is a brief step change in the noise floor as it switches between the converters and back again... much as you might expect.

And its true... it can get caught out in extremis... but I've never had a problem with it in any of my more typical content recordings.

That said, I only use 32bit FP when leaving the machine unattended for dawn chorus recordings and the like. All the choral and orchestral stuff I've recorded on it has been at 24bit fixed point because I've been able to optimise the gain structure properly.

[ghellquist]

What sounds better: 24 bit recording or 32 bit float?

TLTR: it makes no difference. Simply set gains and forget.

Well, there is quite a bit of hype going on here. The market (that is, most of us), seems to expect 32 bit float in all boxes. The reality is a bit more complicated.

Start by remembering that great recordings has been done with 20bit recorders. 20 bit is often the true dynamic range of many so called 24 bit recorders. In use, you simple set the gain low enough to not clip. My suggestion is to aim for -18dBFS in the loud parts. It gives you a whopping 18dB headroom which in reality is totally enormous. The down side is that you get a slight amount more of low level noise - but in a real world situation almost always the room has more noise than that.

A short note: set gain low! Aim for -18dBFS. The advice of aiming for -6dBFS is for the older type of recorders with only 16 bits. In post processing of 24 bit recordiungs you simply normalize. I did a ton of location recordings using a (not quite) 24 bit recorder, and never changed the gain: in my case simple had it set it at 30dB which worked with my microphones and my recorder. From chamber music to rock concerts. Never a clip. No noticeable noise either.

I have since switched to a Zoom F3 and record in 32 bit float. The difference? None! Or to be truthful, maybe slightly lower quality (we are talking miniscule). The reason for the switch was to get a smaller package, nothing else. If we are talking extreme cases, actually the analog inputs of the F3 clips earlier than the ones on my older recorder, but again, it has never happened so far.

[Arpangel]

To me, the bigger the numbers the better, as far as digital goes.

[Hugh Robjohns]

And the manufacturers love you for it!

[Hugh Robjohns]

Start by remembering that great recordings has been done with 20bit recorders.

I think great recordings have been made with far lower 'resolution' than that, proving that the source material is far more important than the mechanism of capture!

20 bit is often the true dynamic range of many so called 24 bit recorders.

This is true for fixed point converters. The best I have measured equates to 21 bits, but most are delivering 19 or 20 bits of meaningful audio data.

The latest 32bit floating point conversion systems which combine high and low gain converters do achieve better dynamic range performance... but in reality it's geared more towards practical convenience than ultimate sound quality.

In use, you simple set the gain low enough to not clip. My suggestion is to aim for -18dBFS in the loud parts. It gives you a whopping 18dB headroom which in reality is totally enormous. The down side is that you get a slight amount more of low level noise - but in a real world situation almost always the room has more noise than that.

In general I agree, although -18dBFS on peaks feels overly generous, especially in situations where you have rehearsals to optimise gain settings.

My personal mantra is to average around -20dBFS and allow peaks no higher than -10dBFS. If it ever hits as high as -6dBFS I'd be getting very squirmy indeed... but if the quieter parts are humming along at -30 or -40dBFS I'm quite happy. As you say ambient room noise is always higher than the digital dither noise with 24bit recording.

[awjoe]

Would you aim for those targets on any and all source material?

[Hugh Robjohns]

Pretty much... although I might give more headroom to something very transient-rich, and less headroom to something mastered or inherently having a small dynamic range. All based on experience, and erring on the side of leaving more headroom rather than less...

But this is for raw recording and tracking, not final mastering.

[awjoe]