Topic: Deity PR-2 Pocket Recorder

[grawk]

I tend to disagree… Consider very very quiet parts in your recording. When using 16 or 24 bit linear storage, you’d be using just a handful of these 16 or 24 bits, so you’ll have relatively big quantization error compared to loud samples. Then, if you’d amplify this quiet audio, you will amplify the quantization error.
In case of 32bfp, even very quiet samples will use plenty of the available bits, so much less quantization errors compared to linear storage. So quiet samples can be amplified without amplifying the quantization error! That’s what’s being shown in the graphs…

Do you have any examples of this happening at live music volume levels?

[TheJez]

I tend to disagree… Consider very very quiet parts in your recording. When using 16 or 24 bit linear storage, you’d be using just a handful of these 16 or 24 bits, so you’ll have relatively big quantization error compared to loud samples. Then, if you’d amplify this quiet audio, you will amplify the quantization error.
In case of 32bfp, even very quiet samples will use plenty of the available bits, so much less quantization errors compared to linear storage. So quiet samples can be amplified without amplifying the quantization error! That’s what’s being shown in the graphs…

Do you have any examples of this happening at live music volume levels?

Unfortunately I’ve had to deal with recordings with both very loud and very quiet songs, and on top of that a very conservative record level because of the unpredictable nature of the performance. So yes, a lot of dynamic processing was needed to make it a listenable recording. Likely the quantization errors are negligible compared to the noise in the quiet parts, but that’s not the point here. I was just trying to explain what was shown in the graph. I think it’s nice not having to worry about amplifying quantization errors during post processing, and I do regard that as an additional benefit of floating point storage, besides the larger dynamic range.
And even loud recordings will have sampled close to 0. It would be nice if these were stored with the same accuracy as high value samples.

[grawk]

I’d just like to see a sample that demonstrates what you’re describing.  I can see how trying to remove the dynamics from a recording can result in errors, however.

[aaronji]

The dynamic range of most shows probably doesn't exceed 60 dB. Maybe 80 dB if you're lucky. Even with a conservative level setting, 24-bit (or 16-bit for that matter) should cover it...

[grawk]

Some amazing sounding recordings were made in the 80s using 14 bit converters and Betamax video recorders.

I still think this argument isn’t really relevant for the pr2 because the limitation is the analog stage not the digital stage. When it’s the right recorder and microphone combination it works and sounds great. If you overload the input or have a really super quiet source it might not be the right device, but it’s not because of 24 or 32 bits.

[TheJez]

Some amazing sounding recordings were made in the 80s using 14 bit converters and Betamax video recorders.

I still think this argument isn’t really relevant for the pr2 because the limitation is the analog stage not the digital stage. When it’s the right recorder and microphone combination it works and sounds great. If you overload the input or have a really super quiet source it might not be the right device, but it’s not because of 24 or 32 bits.

I don’t think there is even an argument   . Somebody posted a graph trying to explain something, somebody else misunderstood the graph, I tried to explain what the graph is trying to show.
No need to do the 24bit vs 32bfp discussion here. Plenty of other threads for that!

[Rairun]

The dynamic range of most shows probably doesn't exceed 60 dB. Maybe 80 dB if you're lucky. Even with a conservative level setting, 24-bit (or 16-bit for that matter) should cover it...

Yeah, you can run a experiment like this:

- Open Audacity and record 10 seconds of audio with your microphone.
- Select all of it and click Edit > Remove Special > Silence Audio
- Export 3 files: 16-bit, 24-bit and 32-bit float.
- Open each file and start adding gain.

This is what I found:

16-bit

0 dB: silence
+10 dB: silence
+15 dB: barely perceptible noise
+20 dB: perceptible noise
+71.224 dB: quantization errors normalised to 0 dBFS

24-bit

0 dB: silence
+50 dB: silence
+60 dB: noise on the edge of perception
+65dB: barely perceptible noise
+70 dB: perceptible noise
+118.474 dB: quantization errors normalised to 0 dBFS

32-bit float

You can amplify it by any number and there will be no quantization errors

Now, if I take one of the quietest passages of music I've ever recorded (a very quiet piano intro), it was around 42 dB quieter than the loudest peaks in the recording (applause at -3 dBFS). If I isolate that part and normalise it to 0 dB, the 24-bit quantization errors will be at least 15 dB quieter than my ears can perceive at full volume with my specific combo of soundcard + headphones - and that's MUCH quieter than the recorder and microphone self-noise, which can be heard with much less digital amplification.

So I agree that quantization errors, if you're using 24-bit files, are negligible. You really need to fuck up your gain staging terribly for it to be relevant. 32-bit float devices are good at reducing noise in practice because they are optimised against their preamp/ADC self-noise. You can much more easily fuck up a recording by bumping against the recorder's analogue self-noise (by feeding too quiet a signal to a preamp/ADC combo that isn't so great) than against the file format's inherent noise (if using 24-bit - it's quite easy to do that with 16-bit).

[Gutbucket]

Just got back to town and following up. If I'm handing my DPA 4061>PR-2 rig off to someone else to run I just leave gain set to 0dB, and set the PR-2 to automatically start recording and then lock itself upon power up.  Those situations are ones in which I know the noise floor of the venue will exceed that of the recording chain regardless of how gain is set.  Just add gain later as needed later. No worries that way. 

In my initial uses, the hot output of DPA Core 4060 was overloading the input stage of the PR-2 even when was set to 0dB gain.  4061 is significantly less sensitive and can accommodate the use of some input gain on the PR-2 without problems in all conditions I've run it so far, but that's not included anything ear-bleedingly loud.

When running the same rig myself I'll add some gain just to get the levels up to something reasonable depending on how loud the material is..  while still leaving plenty of headroom because the meters are VU (averaged) instead of PPM (peak). Makes the resulting recording more usable without processing, especially if playing back directly from the recorder, which is awkward with the PR-2 but I still do it.  Using the same rig to record some acoustic guitar on the back porch a while back I bumped up the gain up significantly. I wish Deity would change the metering from VU (average) to PPM (peak) with a firmware update, or to combo meters, or make them menu switchable.

That's my practical take.  I've not analyzed noise floor or anything yet, but haven't felt the need to as I've only used this rig for shows with high ambient noise floors and have not yet used it for classical recording in a very quiet hall

[Gutbucket]

Beyond the practical take.. (disclaimer, I'm not a digital audio EE!)

I think those staircase waveform images are one of the biggest deciets in digital audio.  First off, a lot of folks, although not so much TS members, seem to believe that staircase represents the shape of the audio output.  But as most here know, the output is always smoothed to a curve.  The question is how accurately the shape of the smoothed output curve reflects the input.  Sampling Theorem tells us that within a given bandwidth, the number of samples needed to do that needs only to exceed that bandwidth by two times in order to be able to retrieve the complete waveform, including all values between those sample times. As long as there are two or more sample points per cycle a steady state sine wave can be fully reconstructed. Extending that, a sinc function of overlapping sine waves is further capable of recreating a complex waveform shape that varies with time.

The devil is in the details of how the waveform is decimated and recreated, the filtering to limit the signal to within the usable bandwith, not in the bit depth and sample rate itself.

It would be alot less misleading if those images showed just the individual sample points at each "step corner" and not the lines connecting them that visually form the stair step representation.  The actual sample points do not get connected by straight lines, they determine a series of overlapping curves that integrate to form the curved output waveform.  The overlapping curves and averaging fit output waveform to the sample points.  There are no stair steps.

It would also be less misleading if such images varied the spacing between sample points along the vertical axis with change of bit-depth, and the spacing between points along the horizontal axis with change of sample-rate.  But such images almost never show that.  Instead they are almost always drawn showing equal spacing along both axes.. as perfectly square stair-shaped steps.

The actual bandwidth limit of a recording is most likely to be solely determined by the dynamic range of the acoustic situation, beyond that by the dynamic range capability of the microphones, and beyond that by the preamp stage or ADC.  A more complicated ADC arrangement designed to switch gracefully between multiple ADCs can extend the ADC range constraint, yet the range of a single 24-bit ADC most likely already exceeds that of the acoustic environment and most microphones.  Within the bandwidth limits determined by the recording chain, the information represented in a 24-bit representation is going to be the same as in a 32-bit float representation.  The 32-bit float storage representation just allows the 24-bit chunk of meaningful data to be shifted up and down as needed in the digital realm, it doesn't provide greater resolution within that meaningful range.

Remember, multiple switching ADCs and 32-bit float representation of the data are two different things.  We can have one without the other, which manufacturers tend to gloss that over.  A multiple switching ADC scheme extends the dynamic range envelope allowing for more lax real-world input gain setting by the user.. or no setting at all.  But in simple terms, the "resolution" within that envelope is defined by the sample rate and determines the high frequency limit of the system, not the reproduction accuracy within the limits of the frequency range.  Higher resolution extends the frequency range, but doesn't increase accuracy within the range.

a 32-bit floating point representation of the output from a non-dithered single 16-bit ADC contains the same information as a 16-bit fixed point representation of it and vice-versa.  The 32-bit floating point digital container itself is vastly larger than the 16-bit digital container, making it capable of representing a far wider dynamic range, but in this case the useful data being output from the single 16-bit ADC and stored inside it remains the same.

Likewise, the electrical noise floor of a 32-bit float recorder is going to be determined by the analog input stage and ADC of the recorder, not by the data storage format.

[dallman]

Honestly I am too impatient to read everything from the last few days, but for what it is worth, for myself, I run the deck only at 24 bit, I have it set for 5v pip, and I have the gain at +15 or +18 depending on the source which is usually rock, jam, reggae, solo, whatever. At 24 bit the gain applied is very important. If I went at 0db gain, I'd have lots of noise. It is no different than any 24 bit recorder in terms of setting gain appropriate to the source, except a preamp is not needed. The recordings at +15 or +18 sound great! I use a Church CA11 mostly and sometimes an AT 853.

[Rairun]

Honestly I am too impatient to read everything from the last few days, but for what it is worth, for myself, I run the deck only at 24 bit, I have it set for 5v pip, and I have the gain at +15 or +18 depending on the source which is usually rock, jam, reggae, solo, whatever. At 24 bit the gain applied is very important. If I went at 0db gain, I'd have lots of noise. It is no different than any 24 bit recorder in terms of setting gain appropriate to the source, except a preamp is not needed. The recordings at +15 or +18 sound great! I use a Church CA11 mostly and sometimes an AT 853.

The gain applied is very important if it is analogue gain. I think it's been said here the PR-2 has no analogue gain, but that sounds weird to me - it'd limit the use of the recorder too much. I'm not saying this is not true, just imho a weird decision.

To use a recorder I'm familiar with as an example, the Roland R-05's converter stage noise is -98 dBFS(A). If it didn't have an internal preamp at all, you'd depend exclusively on the microphone's sensitivity and the loudness of the source to produce a signal as far away from the ADC's self-noise as possible, and the recorder's EIN would be a fixed -98 dB. From experience, using microphones with fairly low sensitivity like the CA-11s, you'd struggle to reach -15 dBFS with zero gain even at the loudest shows (in compliance with UK law). A quiet solo singer-songwriter at a concert hall would very easily be recorded at -50 dBFS with no gain, which when normalised - either in post or by adding digital gain on the recorder - would bring that converter stage noise up to -48 dBFS! Which you can most definitely hear.

The way these recorders attain better EIN is by having analogue gain that increases the signal at a much higher rate than it adds noise (i.e. it might make the signal 10 dB louder while only adding 1 dB of noise). So instead of thinking of the Roland R-05 as a recorder that can reach an EIN of -121.8 dBu(A), I like to think of it as a recorder that gives you clear enough gain to increase the level of the microphone signal so that it can stay as far away from that -98 dBFS(A) noise floor as possible. Or you can use an external preamp for that. Either way, the ADC's noise floor is absolute, you cannot escape it.

Well, you can escape it, but only by having more than one of them calibrated differently, which the PR-2 doesn't have for stereo sources.

My conclusion here - IF it is true that the PR-2 provides no analogue gain at all - is that it is not fit for use with low sensitivity microphones for any show that isn't at least a moderately loud rock concert. Having to rely on the sensitivity of the mics alone doesn't give you a lot of flexibility unless you add a preamp to the chain, which then cancels out the advantage of it providing 5V PiP.

[tapeheadtoo]

So other recorders we're more familiar with, such as Sony M10, A10 etc, apply analog gain whereas the PR-2 only applies digital gain?  What would be the rationale for doing so?

[adrianf74]

The reason to run gain 0 is the gain is all digital so there’s no benefit to running higher than that.

Not that I wasn't to crap all over this, but is there anything documented somewhere showing this or is this based strictly on experience?  Just picked a couple of them up not long ago but haven't opened them up yet to try myself.

[grawk]

I thought I’ve read it but I’m not sure where. it’s what my testing suggests. There’s probably an analog gain stage but the slider at least appears to only change the digital.

[TheJez]

Beyond the practical take.. (disclaimer, I'm not a digital audio EE!)

I think those staircase waveform images are one of the biggest deciets in digital audio.  First off, a lot of folks, although not so much TS members, seem to believe that staircase represents the shape of the audio output.  But as most here know, the output is always smoothed to a curve.  The question is how accurately the shape of the smoothed output curve reflects the input.  Sampling Theorem tells us that within a given bandwidth, the number of samples needed to do that needs only to exceed that bandwidth by two times in order to be able to retrieve the complete waveform, including all values between those sample times. As long as there are two or more sample points per cycle a steady state sine wave can be fully reconstructed. Extending that, a sinc function of overlapping sine waves is further capable of recreating a complex waveform shape that varies with time.

The devil is in the details of how the waveform is decimated and recreated, the filtering to limit the signal to within the usable bandwith, not in the bit depth and sample rate itself.

It would be alot less misleading if those images showed just the individual sample points at each "step corner" and not the lines connecting them that visually form the stair step representation.  The actual sample points do not get connected by straight lines, they determine a series of overlapping curves that integrate to form the curved output waveform.  The overlapping curves and averaging fit output waveform to the sample points.  There are no stair steps.

It would also be less misleading if such images varied the spacing between sample points along the vertical axis with change of bit-depth, and the spacing between points along the horizontal axis with change of sample-rate.  But such images almost never show that.  Instead they are almost always drawn showing equal spacing along both axes.. as perfectly square stair-shaped steps.

The actual bandwidth limit of a recording is most likely to be solely determined by the dynamic range of the acoustic situation, beyond that by the dynamic range capability of the microphones, and beyond that by the preamp stage or ADC.  A more complicated ADC arrangement designed to switch gracefully between multiple ADCs can extend the ADC range constraint, yet the range of a single 24-bit ADC most likely already exceeds that of the acoustic environment and most microphones.  Within the bandwidth limits determined by the recording chain, the information represented in a 24-bit representation is going to be the same as in a 32-bit float representation.  The 32-bit float storage representation just allows the 24-bit chunk of meaningful data to be shifted up and down as needed in the digital realm, it doesn't provide greater resolution within that meaningful range.

Remember, multiple switching ADCs and 32-bit float representation of the data are two different things.  We can have one without the other, which manufacturers tend to gloss that over.  A multiple switching ADC scheme extends the dynamic range envelope allowing for more lax real-world input gain setting by the user.. or no setting at all.  But in simple terms, the "resolution" within that envelope is defined by the sample rate and determines the high frequency limit of the system, not the reproduction accuracy within the limits of the frequency range.  Higher resolution extends the frequency range, but doesn't increase accuracy within the range.

a 32-bit floating point representation of the output from a non-dithered single 16-bit ADC contains the same information as a 16-bit fixed point representation of it and vice-versa.  The 32-bit floating point digital container itself is vastly larger than the 16-bit digital container, making it capable of representing a far wider dynamic range, but in this case the useful data being output from the single 16-bit ADC and stored inside it remains the same.

Likewise, the electrical noise floor of a 32-bit float recorder is going to be determined by the analog input stage and ADC of the recorder, not by the data storage format.

Thanks Gutbucket for your elaborated take on this. Things like resolution and accuracy can be confusing in the (digital) audio world.
IMHO, the practical benefit of the 32bfp format may be limited, except when 'whatever signal processing' is being done which could make the samples go over 0dB, in which case it's essential!
However, again IMHO, I think that if any digital signal processing is done on the output of one or more ADC's in a recorder (either amplification, filtering, dynamics processing, limiting, combining multiple ADC outputs, whatever), information/accuracy will be lost when storing in 24bits linear. This is particularly true for samples close to 0, where just the lowest couple of bits of the sample contain actual information if 16/24 linear storage would be used. Whether this is pure theoretical or practical: For my peace of mind, and if my recorder supports it, I would use 32bfp storage. (Unless that would turn the device into a mono recorder, as is unfortunately the case for the Deity PR-2  ) Personally I don't mind the 33% extra file size.

Apart from all this, I do see mixed reports here about the gain behavior of the Deity... Some say they get noisy recordings when gain set to 0, others say the gain is fully digital, so there is no need to set the gain to anything other than 0... I guess it would be good to get that sorted out...