- Joined
- Feb 23, 2016
- Messages
- 12,839
- Likes
- 19,507
That's a different recorder.
-
WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required as is 20 years of participation in forums (not all true). Come here to have fun, be ready to be teased and not take online life too seriously. We now measure and review equipment for free! Click here for details.
Zoom F6 Portable Field Recorder Review
- Thread starter amirm
- Start date
Yes I am talking about Sony PCM D-100, not Zoom or Sound Devices products. Just to illustrate the dual gain effect in action. It doesn't mean Zoom or Sound Devices is using the same gain stage strategy.
As I mentioned floating point is useful in a situation like this:
...or within the digital processing in a software player before sending to the DAC, the simplest example is lossy codec decoders which the decoded peaks can beyond 0dBFS. If you are doing it with integer, then you need to instruct the decoder to decode at a lower volume, not after decoding, otherwise it will clip.
- Joined
- Apr 24, 2019
- Messages
- 1,503
- Likes
- 2,783
32 bit FP has 23 or 24 bits of resolution due to the length of the mantissa, the patent claims 28 bits of resolution. a 32 bit FP representation can not possibly resolve 28 bits of resolution over its full scale dynamic range. There is some BS here.
- Joined
- Feb 23, 2016
- Messages
- 12,839
- Likes
- 19,507
I thought 32 bit float was able to encode to 25 bits resolution, but yes 28 bit won't be possible. Nor do I expect they'll get 25 bits of dynamic range.
32-bit float has 25 bits of precision, except for subnormal numbers where all of the exponent bits are zero, in this case it works in the same way as integer, the smaller the value, the lower the precision. For example -801dBFS (same formula as dBV to Vrms) is not rounded to zero if encoded in subnormal values, it just loses precision as all exponent bits are used up so you need to consume the significand bits ("mantissa") to represent this value.
32-bit integer can achieve 32 bits of precision only when the values are large enough, in terms of audio , above -6.0206dBFS. It should be emphasized that these are sample values. High volume audio can, and often contain small sample values, like this:
So integer loses precision to small values, float loses precision by consuming exponent bits to maintain precison over a large range.
As a proof that 32-bit float has one more bit than 24-bit in integer domain, try to convert the attached 32-bit float file to 24-bit with any method you like (dither/truncate/noise shaping etc), then compare the converted files in Deltawave.
Anyway, when analog and physical factors are involved, all of the above are irrelevant. Manufacturers who claim "resolution" and such are just delibrately dishonest, not to mention potential artifacts when combining different ADCs, regardless of integer or floating point math.
32-bit integer can achieve 32 bits of precision only when the values are large enough, in terms of audio , above -6.0206dBFS. It should be emphasized that these are sample values. High volume audio can, and often contain small sample values, like this:
So integer loses precision to small values, float loses precision by consuming exponent bits to maintain precison over a large range.
As a proof that 32-bit float has one more bit than 24-bit in integer domain, try to convert the attached 32-bit float file to 24-bit with any method you like (dither/truncate/noise shaping etc), then compare the converted files in Deltawave.
Anyway, when analog and physical factors are involved, all of the above are irrelevant. Manufacturers who claim "resolution" and such are just delibrately dishonest, not to mention potential artifacts when combining different ADCs, regardless of integer or floating point math.
- Joined
- Apr 24, 2019
- Messages
- 1,503
- Likes
- 2,783
Sorry lost a bit, It's been a while. IIRC a full scale sine wave generated with the -lm gcc library in 32bit float vs 24 bit integer is not quite 6dB better for THD, but I might mis-remember. The basic point remains, piecing together A/D's does not buy you the claimed results.
EDIT - No one has mentioned the issue of dithering in these eccentric A/D's.
Last edited:
I mentioned dithering in Ian Shepherd's Youtube video in this post:
https://www.audiosciencereview.com/...table-field-recorder-review.15668/post-505723
jerryfreak
Member
got the zoom back from @Amir yesterday, ran two sets of tests on it from DACs of similar class, the mytek brooklyn DAC+ and the benchmark DAC2
I measured and plotted the Benchmark AD2402-96 (AD2K) as a reference. Somewhat dated vintage gear but a solid performer that can take a hot input with great performance
yeah that low frequency distortion is unshakable at any input level
is -70 to -80 dB distortion below 50Hz a problem? not sure, its pretty low on the audibility curve of human hearing
knowing what i know now, will i be using the HPF? you bet
at the end of the day it seems like the sound quality of a handheld (at best) in a prosumer body.
I measured and plotted the Benchmark AD2402-96 (AD2K) as a reference. Somewhat dated vintage gear but a solid performer that can take a hot input with great performance
yeah that low frequency distortion is unshakable at any input level
is -70 to -80 dB distortion below 50Hz a problem? not sure, its pretty low on the audibility curve of human hearing
knowing what i know now, will i be using the HPF? you bet
at the end of the day it seems like the sound quality of a handheld (at best) in a prosumer body.
- Joined
- Feb 23, 2016
- Messages
- 12,839
- Likes
- 19,507
So from these results it appears the low frequency noise floor is heavily modulated by the signal level in the F6. And at higher signal levels distortion is somewhat high.
Was this done by recording results of the F6 to the SD card or was it a loop-thru? And did you do any of these with the 32 bit float mode of the F6?
jerryfreak
Member
no, as they exaggerate the max input level. it cant take over +17 dBU cleanly, and increasing input from +4 to +17 added little. There are tests at +22 and +23 dB there for you to evaluate. all zoom tests were done at 32bit float
there were some other preliminary tests in 24 bit. 32float added a few dB to the noise floor but had no effect on the high distortion levels
https://taperssection.com/index.php?topic=190161.msg2339748#msg2339748
Last edited:
AnalogSteph
Major Contributor
I bet they're using some sort of low power opamp in there that's really struggling at high levels.
The AD2K seems to be using a CS5396? 7th order (!) tri-level delta sigma, no wonder its noise floor looks like idle tone city. Stability of such a high order modulator could only ever be rather conditional. Looks like Crystal had AKM beat in late '97, whose otherwise similar AK5392 came out two months later, with a measly 116 dB(A) rather than 120 dB(A) dynamic range and still lacking 96 kHz support (which was only available with the AK5393 in early 1999).
I do wonder what was going on inside the Prism Sound AD-2... you don't see a typical dynamic range of 131.5 dB unweighted every day (even if at a very high +28 dBu in - at +18 dBu you have to make do with 128 dB), let alone in something released in 1998! No wonder these guys got into test equipment later. There's also digitally selectable analog input gain in 0.5 dB steps from +5 dBu to +28 dBu. (In the above Benchmark you had to be content with +14 dBu to +24 dBu in 2 dB steps.)
Now you had me going down the rabbit hole of vintage ADCs... again.
The AD2K seems to be using a CS5396? 7th order (!) tri-level delta sigma, no wonder its noise floor looks like idle tone city. Stability of such a high order modulator could only ever be rather conditional. Looks like Crystal had AKM beat in late '97, whose otherwise similar AK5392 came out two months later, with a measly 116 dB(A) rather than 120 dB(A) dynamic range and still lacking 96 kHz support (which was only available with the AK5393 in early 1999).I do wonder what was going on inside the Prism Sound AD-2... you don't see a typical dynamic range of 131.5 dB unweighted every day (even if at a very high +28 dBu in - at +18 dBu you have to make do with 128 dB), let alone in something released in 1998! No wonder these guys got into test equipment later. There's also digitally selectable analog input gain in 0.5 dB steps from +5 dBu to +28 dBu. (In the above Benchmark you had to be content with +14 dBu to +24 dBu in 2 dB steps.)
Sorry for another verbose post but there are something to concern when testing similar devices in future. RMAA puts a limit of 133dB displayed DNR (slightly higher for higher sample rates and bit depths). For noise levels, RMAA either adds dither in integer reference files or some very low level noise (below -200dB) in float reference files. Digital silence is displayed as -400dB instead of -inf or values below -758dB. If your test results have -400dB noise level it means your digital signal chain truncated the noise and it is not ideal. Ideally noise level should be identical or similar to the reference file.
As mentioned before many lossy codecs are floating point based, many encoders accept 24-bit input and many decoders support 32-bit float decoding. For example I attached an 140kbps mp3 and it can preserve the DNR of the original 24-bit reference file. It is not instantaneous DNR, but RMAA doesn't know about it.
The decoded mp3 waveform is not deformed as well, at least for a screenshot with several hundred pixels of resolution.
There is a bit tricky to test these ADCs, perhaps we need to use test signals with large and sudden level change and inspect the transitional behavior, as well as test signals with different fading curves, in single and multi-tone, and unfiltered but alias-free square wave (e.g. 750Hz square wave at 48kHz sample rate).
As mentioned before many lossy codecs are floating point based, many encoders accept 24-bit input and many decoders support 32-bit float decoding. For example I attached an 140kbps mp3 and it can preserve the DNR of the original 24-bit reference file. It is not instantaneous DNR, but RMAA doesn't know about it.
The decoded mp3 waveform is not deformed as well, at least for a screenshot with several hundred pixels of resolution.
There is a bit tricky to test these ADCs, perhaps we need to use test signals with large and sudden level change and inspect the transitional behavior, as well as test signals with different fading curves, in single and multi-tone, and unfiltered but alias-free square wave (e.g. 750Hz square wave at 48kHz sample rate).
Last edited:
jerryfreak
Member
Zoom engineers suspect my unit is defective and are sending me a replacement. will test when i receive
- Joined
- Feb 23, 2016
- Messages
- 12,839
- Likes
- 19,507
You having it sent to Amir too?
jerryfreak
Member
no, i will be testing it here
This is an interesting thread discussing floating point with SNR. I've had many of these discussions before and it is tricky to navigate, somewhat because of terminology. You have to consider separating the concept of traditional converter quantization noise from what is going on numerically. What these companies are attempting to do is to construct a numerically higher SNR signal from two quantized signals. To Scott's point, that is extremely difficult to do without adding more noise or distortion, but ignore that for a moment.
Back to numerics... A floating point number maintains its full 24 bit mantissa precision at any value within the exponent range. In a digitally created signal (no real life recorded noise), you can create the signal with a peak level of 0.001, or a peak level of 1.000, or a peak level of 1000.0 and they have exactly the same SNR, -144dB. The way we measure SNR, using a steady state sine tone of 0dBFS is not going to show any advantage for floating point. It should measure identically to 24bit fixed. In truth, the "quantization noise floor" is actually modulating with the signal values, with the highest quantization noise at the peaks of the sine tone and the quantization noise reducing as the values approach 0. However, this isn't really an advantage with the way we hear or with the way we measure. The advantage is that real signals have quiet sections and louder sections, and you can compress the signal, or turn up the quieter sections, without turning up the quantization noise there. So, the goal of these field recorders is to essentially use two different preamp gains to capture quieter signals with higher gain and louder signals with lower gain, and then try to combine them into one signal. Since the gains are different, you need to go to higher precision numbers to reconstruct the combined range. Calling this floating point is somewhat misleading as I doubt the difference in preamp gains is more than 20-30dB. At best, it is similar to 30 bit fixed point audio, maybe as high as 180dB numerical representation.
But back to that difficult to do piece... Let's say you have a quiet section of the music, you really just want to use the quiet converter as adding in the loud converter will just add noise into your signal with not much value. Then the music gets loud and you want to use that converter, the switch to that converter is very difficult to do without adding noise or distortion during the switch. You get the idea... So the idea that you would switch converters during the middle of every single cycle of a sine tone during an SNR measurement seems like a worst case scenario. Ultimately, you would be better off just sticking with the loud converter as that will dictate your quantization noise based SNR.
I think it's more useful to think of techniques like this as an attempt to save a poorly recorded, or challenging to record, signal than a way to increase SNR. It's just a way of recording the signal with two different preamp settings and merging them conveniently into a single file.
Back to numerics... A floating point number maintains its full 24 bit mantissa precision at any value within the exponent range. In a digitally created signal (no real life recorded noise), you can create the signal with a peak level of 0.001, or a peak level of 1.000, or a peak level of 1000.0 and they have exactly the same SNR, -144dB. The way we measure SNR, using a steady state sine tone of 0dBFS is not going to show any advantage for floating point. It should measure identically to 24bit fixed. In truth, the "quantization noise floor" is actually modulating with the signal values, with the highest quantization noise at the peaks of the sine tone and the quantization noise reducing as the values approach 0. However, this isn't really an advantage with the way we hear or with the way we measure. The advantage is that real signals have quiet sections and louder sections, and you can compress the signal, or turn up the quieter sections, without turning up the quantization noise there. So, the goal of these field recorders is to essentially use two different preamp gains to capture quieter signals with higher gain and louder signals with lower gain, and then try to combine them into one signal. Since the gains are different, you need to go to higher precision numbers to reconstruct the combined range. Calling this floating point is somewhat misleading as I doubt the difference in preamp gains is more than 20-30dB. At best, it is similar to 30 bit fixed point audio, maybe as high as 180dB numerical representation.
But back to that difficult to do piece... Let's say you have a quiet section of the music, you really just want to use the quiet converter as adding in the loud converter will just add noise into your signal with not much value. Then the music gets loud and you want to use that converter, the switch to that converter is very difficult to do without adding noise or distortion during the switch. You get the idea... So the idea that you would switch converters during the middle of every single cycle of a sine tone during an SNR measurement seems like a worst case scenario. Ultimately, you would be better off just sticking with the loud converter as that will dictate your quantization noise based SNR.
I think it's more useful to think of techniques like this as an attempt to save a poorly recorded, or challenging to record, signal than a way to increase SNR. It's just a way of recording the signal with two different preamp settings and merging them conveniently into a single file.
jerryfreak
Member
mfr is active in addressing the distortion issue... will post updates when i can
It's already mentioned many many times, even on the same page, that 32-bit float has 25 bits of precision, not 24.
https://www.audiosciencereview.com/...depth-should-i-set-my-dac-to.8956/post-255104
https://www.audiosciencereview.com/...table-field-recorder-review.15668/post-507610
I think you're basically right. Although, in practice, it might be more fair to say that floating point has 144dB of dynamic range and 24 bit fixed has 138dB of dynamic range (23 bit equivalent), due mainly to the fact that the negative swinging side of the audio signal doesn't add to Dynamic Range. Haven't thought too much about that detail though, so I could be wrong.
The point of my post was more about the issue of relating precision to SNR in a floating point scheme.
Similar threads
