Focusrite 18i20 (gen.2), Lavry DA10, REW interpretation

[Blumlein 88]

Using 8 sweeps reduces the contribution of noise.

 

[Blumlein 88]

This is using 8 sweeps with larger FFTs. Blue is 1 m FFT and red is 4 m FFT. Last couple posted were 256 k FFTs.

 

[HenryL]

(@Blumlein 88 ) Your measurements are pretty consistent with my findings it seems. I experimented with averaging of multiple traces and also found it reduced the random noise element, as did a larger/longer sweep which seemed better. In the end I stuck to a single sweep with a low (-est, or nearly lowest) available size/time for the sample rate because I was interest in seeing where the noise showed up as much as the overall curve trend.

I've tried to collect a representative selection of all my previous measurements together into one overview picture. It's a bit much for one picture but here it is as two pictures, the second one zooming in a little for a bit more detail at the lower 44.1/48 end. It's a bit busy obviously but the patterns are fairly easy to identify. I didn't do anything at 176.4kHz hence it jumps up to to 192. I'll add a little clearer exposition of what's what here shortly - in the meantime duty calls...

So here's an overview of most of the the top end of all of it together:

And here's a slightly better view of the 44.1 and 48 zone, with graph-spacing switched on:

 

[Blumlein 88]

Here is 44.1 khz at a high level. I'm not seeing the amount of ripple you show. There is maybe a tiny hint of it, but I'm much more zoomed in than your graphs. Maybe the filtering is different in the later 192 khz capable units. Even though the same chips are used.

 

[Blumlein 88]

I sometimes record sweeps and view them in a spectrogram view in Audacity. This lets you see harmonic and intermodulation distortion, aliasing/imaging, and if bad enough jitter artifacts.

When I do that you can adjust the rate of the sweep vs the size of the FFT to minimize the effects of noise. You essentially are doing that with multiple sweeps and larger FFTs in REW. If you could lengthen the sweeps you could do it with smaller FFTs.

Here is an example where I did a sweep to 96 khz while recording at 48 khz. In this case the spectrogram was set to be light gray at -100 dbFS. You can change that and sometimes find other things. For instance the 18i20 has a series of tones spaced every 1000 hz all the way to the maximum frequency. These are down around -126 dbFS however.

 

[HenryL]

Maybe the filtering is different in the later 192 khz capable units. Even though the same chips are used.

Yes this seems possible, different options selected on the chips perhaps - though another variable is the choice of output channel. I only used outputs 8 and 3 and inputs 8 and 3. I noticed no difference between inputs 8 and 3 but between outputs 8 and 3 there was a clear difference in the behaviour with low level sweeps in that the output was much more wiggly, whilst at high level they looked very similar (shorter wiggles). You used a main output presumably meaning output 1 or 2. There are maybe implementation differences with those outputs - I think they may be on the 2 channel DAC-only chip whilst the others are all on combined DAC/ADC ships. I think they have soft-start muting etc (though that seems unlikely to be a factor here).

 

[HenryL]

I sometimes record sweeps and view them in a spectrogram view in Audacity. This lets you see harmonic and intermodulation distortion, aliasing/imaging, and if bad enough jitter artifacts.

When I do that you can adjust the rate of the sweep vs the size of the FFT to minimize the effects of noise. You essentially are doing that with multiple sweeps and larger FFTs in REW. If you could lengthen the sweeps you could do it with smaller FFTs.

Here is an example where I did a sweep to 96 khz while recording at 48 khz. In this case the spectrogram was set to be light gray at -100 dbFS. You can change that and sometimes find other things. For instance the 18i20 has a series of tones spaced every 1000 hz all the way to the maximum frequency. These are down around -126 dbFS however.

Interesting. The horizontal axis is presumably time over which the sweep was performed, ie about 48 seconds overall? And relatively strong aliasing from frequencies 24-26kHz ?

 

[HenryL]

For comparison with the above here's my collected Lavry DAC -> 18i20 line input 3 plots. (Max 96kHz on the Lav)

 

[HenryL]

So to round up my own observations so far, based on high level REW sweeps at -2dBFS, low level sweeps at -40dBFS with the input gain adjusted in each case to give a REW 'check level' reading for the measurement of around -12dBFS (plus or minus half a dB or so):

- The behaviour of the 18i20 at low levels varies, mainly depending on the choice of output. (Changing the input didn't show up any differences with a given 18i20 output and whilst using the lavry output to the two different 18i20 inputs there were not the noise blooms evident with the 18i20 output regardless of which input was chosen)

- At 44.1 kHz the 18i20 is prone to wiggles in the FR plot extending down to ~16kHz at high sweep levels, whilst at low sweep levels output 8 was smooth whilst output 3 was wiggly. Maybe a magnitude of 0.05 dB with a high level signal and 0.2dB with a low level signal on output 3. Don't know what if any audible effect this may have.

- At 48kHz the 18i20 high sweep levels now stop wiggling, but the low level sweeps are the same as at 44.1 with output 3 being wiggly whilst output 8 is smooth.

- From 88.2kHz we see a quite different rolloff trajectory. There isn't really anything I'd call a 'rolloff' visible at 44.1 and 48kHz, but from 88.2 suddenly there is.

- With 88.2 The high level sweeps are smooth regardless of which output used but there is much more noise at the knee with the low level sweeps, higher at 88.2 even than at 96 for some reason. Both outputs show the same HF noisiness with low level sweeps, but once again output 3 has wiggles.

- At 96kHz I only took measurements from output 8 (the non-wiggly one) so I don't know if low level sweeps on output three would still have been wiggly at this sample
rate (but probably). 96kHz is less noisy than 88.2.

- At 192kHz the HF noise explodes with the low-level signal. The high level signal is significantly better but with noise increasing gradually over a greater HF sweep range than at lower sample rates,
(

(It is common to just omit part of the filter chain at higher sample rates, though in more modern parts this may only occur past 96 kHz.

?)

 

[Blumlein 88]

Interesting. The horizontal axis is presumably time over which the sweep was performed, ie about 48 seconds overall? And relatively strong aliasing from frequencies 24-26kHz ?

Yes.

 

[Blumlein 88]

I tried 44.1 khz on output 3,6, and 8. 256 k FFT. -40 dbFS output made up by gain on input. 8 sweeps, All looked the same.

Output 8 at -40 dbFS, 88.2 khz, 256k FFT 8 sweeps.

 

[Blumlein 88]

Here are the 1 khz spurious tones in the 18i20. Lower level than I remembered.

 

[HenryL]

@Blumlein 88 - Thanks for running those. Intriguing that you don't get any of the things I did at 44.1 though your 88.2 looks more similar to my result.

One small aspect: when you say you did 8 sweeps do you mean you did 8 separate sweeps and selected 'average the responses', or selected 8 as the number of repetitions in the measurement panel? I'm guessing the latter. I'll try it the same way as you did it as a sanity check.

The x8 will reduce the visibility of random noise compared to my single sweeps. If the maths is the same as increasing the sweep time (seems intuitively reasonable...) then it should have the effect of 3dB noise reduction for each doubling, so your plots should look 9dB less noisy than mine. The REW guidance said that to reduce noise effects it was generally better to use a single sweep but to make it longer, rather than to average repeated sweeps - but I forget the reason for the moment.

Anyway I don't think the low level wiggling is going to behave like random noise so it wouldn't account for that difference, not obviously anyway, and there is evidently an implementation difference - or faulty component perhaps - in my device to account for it varying consistently between outputs.

Anyway - please don't run any more tests on my account, ie. not unless you have your own interest in this too! It all takes time...

 

[HenryL]

Here are the 1 khz spurious tones in the 18i20. Lower level than I remembered.

View attachment 308282

How was this taken? Looks like REW RTA.. Is this just listening to an unconnected input, or is it listening to an output looped back to an input? And at what input gain level?

cheers/Henry

 

[Blumlein 88]

@Blumlein 88 - Thanks for running those. Intriguing that you don't get any of the things I did at 44.1 though your 88.2 looks more similar to my result.

One small aspect: when you say you did 8 sweeps do you mean you did 8 separate sweeps and selected 'average the responses', or selected 8 as the number of repetitions in the measurement panel? I'm guessing the latter. I'll try it the same way as you did it as a sanity check.

The x8 will reduce the visibility of random noise compared to my single sweeps. If the maths is the same as increasing the sweep time (seems intuitively reasonable...) then it should have the effect of 3dB noise reduction for each doubling, so your plots should look 9dB less noisy than mine. The REW guidance said that to reduce noise effects it was generally better to use a single sweep but to make it longer, rather than to average repeated sweeps - but I forget the reason for the moment.

Anyway I don't think the low level wiggling is going to behave like random noise so it wouldn't account for that difference, not obviously anyway, and there is evidently an implementation difference - or faulty component perhaps - in my device to account for it varying consistently between outputs.

Anyway - please don't run any more tests on my account, ie. not unless you have your own interest in this too! It all takes time...

Yes I selected doing 8 sweeps in the menu. And no FR wiggling is not effected by that. I do have another interface with rippled responses. It happens at all levels.

 

[Blumlein 88]

How was this taken? Looks like REW RTA.. Is this just listening to an unconnected input, or is it listening to an output looped back to an input? And at what input gain level?

cheers/Henry

I had output 8 connected to input 1, but no signal going out. So listening using the RTA function. I've checked it with signal and it's still there.

 

[JohnPM]

I am also unaware of REW's sound API limitations on the Mac. You may very well need to keep its sample rate sync'd with system settings. Not sure whether the same limit of 16 bit samples applies when using Java I/O, which is the case on Windows.

On macOS REW uses the sample format specified in the Mac's Audio Midi Setup utility for the interfaces being used. There is no 16-bit limitation. REW cannot control the interface sample rate on the Mac so the rate in Audio Midi Setup should match the rate in REW otherwise the OS will resample.

On Windows the JavaSound drivers are limited to stereo and 16-bit data and REW cannot control the sample rate. That is not the case for the WASAPI exclusive Java drivers (device entries whose names start EXCL) or ASIO drivers, for both of those REW uses the highest sample resolution the interface supports and configures the sample rate to match the rate chosen in REW.

The actual sample resolutions REW is using are shown at the bottom of the main window, along with a bit activity display for input data. Here is an example where the sample formats for input and output are both 32-bit but REW has detected that the input data is actually 24-bit, the bottom 8 bits are always zero.

 

[HenryL]

On macOS REW uses the sample format specified in the Mac's Audio Midi Setup utility for the interfaces being used. There is no 16-bit limitation. REW cannot control the interface sample rate on the Mac so the rate in Audio Midi Setup should match the rate in REW otherwise the OS will resample.

On Windows the JavaSound drivers are limited to stereo and 16-bit data and REW cannot control the sample rate. That is not the case for the WASAPI exclusive Java drivers (device entries whose names start EXCL) or ASIO drivers, for both of those REW uses the highest sample resolution the interface supports and configures the sample rate to match the rate chosen in REW.

The actual sample resolutions REW is using are shown at the bottom of the main window, along with a bit activity display for input data. Here is an example where the sample formats for input and output are both 32-bit but REW has detected that the input data is actually 24-bit, the bottom 8 bits are always zero.
View attachment 308514

Thank you John

 

[HenryL]

I sometimes record sweeps and view them in a spectrogram view in Audacity. This lets you see harmonic and intermodulation distortion, aliasing/imaging, and if bad enough jitter artifacts.

When I do that you can adjust the rate of the sweep vs the size of the FFT to minimize the effects of noise. You essentially are doing that with multiple sweeps and larger FFTs in REW. If you could lengthen the sweeps you could do it with smaller FFTs.

Here is an example where I did a sweep to 96 khz while recording at 48 khz. In this case the spectrogram was set to be light gray at -100 dbFS. You can change that and sometimes find other things. For instance the 18i20 has a series of tones spaced every 1000 hz all the way to the maximum frequency. These are down around -126 dbFS however.

@Blumlein 88 How did you make these measurements by the way? I'm interested in doing something similar for a couple of ADC's I have as well as the 2nd Gen 18i20.

I take it this is using a TC Impact Twin as the DAC (I've had two of these in the past, 2nd replacing the first under retailers warranty, then the 2nd died in a similar time...).

Because of the alignment of the graphs it looks like the sweep was recorded along with spectrum, so how did you have two different sample rates running at the same time in the same system? Or was the test sweep recorded separately and aligned 'manually' here? (Recorded as a pre-run on the measurement machine at the higher sample rate?)

I've revived an old iMac which has an optical line out capability with which I can drive an 18-yr-old RME ADI2 as the sweep signal source. I guess this should be useable for the above kind of experiments. The output doesn't seem to go above 96kHz unfortunately so limited to sweeps up to 48kHz though this would seem enough to cover the most likely ultrasonic sources of aliasing in practice.

I haven't seen how to make a similar 'spectrogram over time' plot with REW. I'm not familiar with Audacity and I don't really want to have to familiarise with another piece of software if possible but if need be...

I've tried to come up with other ways of capturing aliasing using REW but none of them have seemed to be workable with REW so far. I thought of using bandwidth limited noise from say 20kHz to 48kHz, or a sweep from 20kHz upwards but the REW generator doesn't seem to allow me to do these. If it had worked then I planned to just compare the spectra up to 20kHz with and without the test ultrasonic signal input to see what if any new audible frequencies put their heads above the parapet.

Or can anyone point me to another / classic / standard way of looking for aliasing, ideally using REW, that I'm missing ?

Cheers/ Henry

I've tried to a few other ideas for

 

[Blumlein 88]

@Blumlein 88 How did you make these measurements by the way? I'm interested in doing something similar for a couple of ADC's I have as well as the 2nd Gen 18i20.

I take it this is using a TC Impact Twin as the DAC (I've had two of these in the past, 2nd replacing the first under retailers warranty, then the 2nd died in a similar time...).

Because of the alignment of the graphs it looks like the sweep was recorded along with spectrum, so how did you have two different sample rates running at the same time in the same system? Or was the test sweep recorded separately and aligned 'manually' here? (Recorded as a pre-run on the measurement machine at the higher sample rate?)

I've revived an old iMac which has an optical line out capability with which I can drive an 18-yr-old RME ADI2 as the sweep signal source. I guess this should be useable for the above kind of experiments. The output doesn't seem to go above 96kHz unfortunately so limited to sweeps up to 48kHz though this would seem enough to cover the most likely ultrasonic sources of aliasing in practice.

I haven't seen how to make a similar 'spectrogram over time' plot with REW. I'm not familiar with Audacity and I don't really want to have to familiarise with another piece of software if possible but if need be...

I've tried to come up with other ways of capturing aliasing using REW but none of them have seemed to be workable with REW so far. I thought of using bandwidth limited noise from say 20kHz to 48kHz, or a sweep from 20kHz upwards but the REW generator doesn't seem to allow me to do these. If it had worked then I planned to just compare the spectra up to 20kHz with and without the test ultrasonic signal input to see what if any new audible frequencies put their heads above the parapet.

Or can anyone point me to another / classic / standard way of looking for aliasing, ideally using REW, that I'm missing ?

Cheers/ Henry

I've tried to a few other ideas for

Yes it was an Impact Twin. Mine is still working I just don't have any Firewire machines for it. I mostly use it in stand alone mode for karaoke parties.

Yes, I used two laptops one for DAC, and one for ADC so they could be different sample rates. Now you could use Multitone as it will allow two devices connected to a single laptop running at different sample rates. Or you can create the files as wav files and play them with a DAC to be analyzed on the ADC capture.

One way I know is using static tones. Use REW in RTA mode while playing a tone above the sample rate limit of the ADC. Run the ADC at say 48 khz, and run the DAC at 96 or 192 khz. Play tones at 25 khz or higher. You could do it every couple of khz or whatever. Record it with the ADC and you'll see as you move the tone higher it will show up at lower frequencies in the ADC. 25 khz tone will show up at 23 khz. 27khz tone will show up as a 21 khz tone when aliased. 30 khz tone will show up as an aliased 18 khz tone in the ADC. The tones will mirror around the 24 khz cut-off of the ADC. That is probably the most straightforward if tedious way with REW. You can compare the level of the aliasing vs frequency.