• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Beta-test: DeltaWave Null Comparison software

Blumlein 88

Grand Contributor
Forum Donor
Joined
Feb 23, 2016
Messages
20,521
Likes
37,050
Looks like a fair size difference in the 30-50 hz range. And linearity is very low.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,626
Likes
10,202
Location
North-East
OK here are three example files. 100.wav is the reference. my Deltawave thinks mpd3 is reversed polarity but I told it not to invert. I see more than 10dB difference in the spectrum of delta. I deleted all the really bad ones but I can capture some more later.
https://www.dropbox.com/s/53spkr9hglakrl3/Dropbox.zip?dl=0

Thanks, looks like there is a clock drift (which is expected when you use two separate units with separate clocks). Because there are a few portions of the file that are almost an exact match (in the frequency domain) to other parts of the file, with just a different amplitude, this causes DW to think it's a simple periodic waveform (measurement waveform prompt). I just tell it that it's not, and it continues fine.

Here's the first file result with clock drift correction for mpd2:

1672433457979.png


Here's the match for mpd3. Note that mpd3 recording started a number of seconds earlier than the reference, so I trimmed two seconds off the front by setting Trim Front to 2. DeltaWave finds the initial delay mismatch if it's within about 2x the FFT size number of samples. If it's more than that, you need to trim one of the files to get them to at most a few second delay between them:
1672434621775.png


And here are the settings I used with both files:

1672434845075.png
 

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
@pma,
With "non-linear correction" for mag/phase I get a -80dB correlated null and PKmetric at -85dB, a fair result.
Phase rotation at LF and HF corner is the main culprit but there is also visible magitude passband ripple and both will be undone with nonlin correction (which also nails nonlinearity).
1672434972541.png


And there is a 0.6dB level difference, quite visible in raw files which probably gave the main cues for your ABX
 

pma

Major Contributor
Joined
Feb 23, 2019
Messages
4,591
Likes
10,727
Location
Prague
And there is a 0.6dB level difference, quite visible in raw files which probably gave the main cues for your ABX
There is no chance to simply fix the level, and the "level difference" is much lower.
Initial peak values Reference: -2.654dB Comparison: -2.438dB
Initial RMS values Reference: -15.264dB Comparison: -15.224dB

Final peak values Reference: -2.654dB Comparison: -2.533dB
Final RMS values Reference: -15.264dB Comparison: -15.322dB

ABX works even with the "matched" file. The biggest difference is the timbre of the guitar in the left ear - try it. It is easy to hear it. Matched waveforms:

1672435760978.png


It does not make sense to make all sorts of corrections, because the files do sound different. The system that measures THD much better than -120dB and SINAD 112dB produces audibly different results. This only shows that the measuring methods we (yes not all of us) consider most important may not tell much and may not correlate with our hearing. We have known that. Only it is not always easy to bring the proof and this is one possible proof.
 

Blumlein 88

Grand Contributor
Forum Donor
Joined
Feb 23, 2016
Messages
20,521
Likes
37,050
There is no chance to simply fix the level, and the "level difference" is much lower.
Initial peak values Reference: -2.654dB Comparison: -2.438dB
Initial RMS values Reference: -15.264dB Comparison: -15.224dB

Final peak values Reference: -2.654dB Comparison: -2.533dB
Final RMS values Reference: -15.264dB Comparison: -15.322dB

ABX works even with the "matched" file. The biggest difference is the timbre of the guitar in the left ear - try it. It is easy to hear it. Matched waveforms:

View attachment 253552

It does not make sense to make all sorts of corrections, because the files do sound different. The system that measures THD much better than -120dB and SINAD 112dB produces audibly different results. This only shows that the measuring methods we (yes not all of us) consider most important may not tell much and may not correlate with our hearing. We have known that. Only it is not always easy to bring the proof and this is one possible proof.
As I said in the other thread, using non-linear correction is in one sense cheating. However it is a tool to find what is causing the difference in null levels. In this case there is a low corner filter and some variable response in the devices between 400hz and 10 khz. Those should be measurable with a good FR test though it might slip thru on just doing 1 khz. It does show overly simplified testing has some blind spots.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,626
Likes
10,202
Location
North-East
Hello Paul, I have come to an interesting problem today. Normally I prepare my files for listening tests through the soundcard which has the same clock for its ADC and DAC and when comparing the loop product to the original music file, Deltawave gives negligibly low error numbers and the files are indistinguishable in the ABX test. Today, I have decided the loop constituted of Topping D10s DAC (connected to PC through USB-ISO to cut the ground loop) and E1DA Cosmos ADC. For THD or IMD measurements, this loop gives excellent numbers. However, in a loop with a music file, and the result compared to the original file (both previously level matched and time aligned in AA), the result is horrible. Not only in pkmetrics and error distribution, but the difference between the original and the loop product is easily audible. Would you have an idea what could be the issue? It is definitely not the Deltawave itself.

View attachment 253509

View attachment 253510

Hi Pavel,

As Klaus said, the issue appears to be large, non-linear timing differences between the original and the loopback recording. DeltaWave can remove a linear clock drift, but it can't fix non-linear variations (including jitter and noise- or signal-modulated clock errors) which will result in large phase differences, as in here:

1672436113386.png



The clock drift computation attempts to straighten out the linear clock differences, but large variations still remain after this (Y axis is the clock error in samples) in the following chart, corrected is the top, white line still showing large variations throughout the file:
1672436492539.png



There are also large frequency errors, which are harder to explain with the two devices you mention, but this is clearly there, around 35-40Hz, and then between 2 and 3khz, which is (unsurprisingly) audible:
1672436396910.png


1672436558796.png


This doesn't seem to perform as expected at all, so make sure the ADC is configured correctly (mono vs stereo mode) and that there really isn't a ground loop of some kind despite your attempts to break it...
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,626
Likes
10,202
Location
North-East
There is no chance to simply fix the level, and the "level difference" is much lower.
Initial peak values Reference: -2.654dB Comparison: -2.438dB
Initial RMS values Reference: -15.264dB Comparison: -15.224dB

Final peak values Reference: -2.654dB Comparison: -2.533dB
Final RMS values Reference: -15.264dB Comparison: -15.322dB

ABX works even with the "matched" file. The biggest difference is the timbre of the guitar in the left ear - try it. It is easy to hear it. Matched waveforms:

View attachment 253552

It does not make sense to make all sorts of corrections, because the files do sound different. The system that measures THD much better than -120dB and SINAD 112dB produces audibly different results. This only shows that the measuring methods we (yes not all of us) consider most important may not tell much and may not correlate with our hearing. We have known that. Only it is not always easy to bring the proof and this is one possible proof.

The timbre differences are fairly obvious from the frequency delta plot, errors of 6-7dB in the most sensitive audible region between 2-3kHz are hard to ignore!
 

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
It does not make sense to make all sorts of corrections, because the files do sound different.
I haven't made a listening test or ABX so far, but will do....after trying to do a better level match, or even better, stepping level differences in 0.1dB steps through range from -0.5dB to +0.5dB. Usually I find systematic "sound signature" differences other than simple level to survive a slight level mismatch. If not, it was just the level difference alone.

And I would agree partly, DW has so many parameters that it's easy to get lost and "overcorrect". With a bit of fiddling I can arrive at -97dB correlated null and -102dB PKmetric.
Residual now is "clean" enough to expose slight dynamic gain drifts (reference voltage instabilities) and clock drift and does not expose much distortion (as expected)
 

Sokel

Master Contributor
Joined
Sep 8, 2021
Messages
5,830
Likes
5,758
Excuse my non-technical but if Pavel's USB isolator is like one I tested (not full speed) when I first using Multitone can mess things up pretty good even if it's not using high bit rates.
Just for consideration.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,626
Likes
10,202
Location
North-East
I haven't made a listening test or ABX so far, but will do....after trying to do a better level match, or even better, stepping level differences in 0.1dB steps through range from -0.5dB to +0.5dB. Usually I find systematic "sound signature" differences other than simple level to survive a slight level mismatch. If not, it was just the level difference alone.

And I would agree partly, DW has so many parameters that it's easy to get lost and "overcorrect". With a bit of fiddling I can arrive at -97dB correlated null and -102dB PKmetric.
Residual now is "clean" enough to expose slight dynamic gain drifts (reference voltage instabilities) and does not expose much distortion (as expected)
It's important to understand the differences between linear and non-linear corrections. As Dennis said, non-linear EQ corrections are "cheating" in a sense that these will correct for errors that might be very audible. They exist in order to help diagnose what's causing the difference. Non-linear phase errors reduced by non-linear EQ tells you that there are timing variations that are not corrected by simple drift removal. These errors can very well be audible. Non-linear level errors are even more likely to be audible, so when non-linear level EQ causes a large decrease in error -- this is likely the cause.

For any real "audibility" testing, I'd leave non-linear EQ settings off, completely.

The one setting that might help improve the error numbers a bit is the Auto-trim at start & end. This should be ON in most cases, as this removes large differences in the lead-in and lead-out of the two files. With this option turned off (@pma had it off) the error result will include the potentially large differences at the start and/or the end of the two files.
 

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
@pkane, I'm still using DW 2.0.2 and I'm getting completely different results (without nonlin correction) than what you showed.... and now I found why:

I've used right channels for analysis (probably as selected from the last test I did) and this behaves much better than the left to left channel comparison where I now get similar results than you and those indicate somethning is broken in the DAC-->ADC chain for the left channels which then also explains the ABX.

Right channel results (nonlin cor. off):
1672439055698.png

There is passband ripple of the ADC and DAC but otherwise a close match.

And it is a linear phase ripple because phase is clean:
1672439182127.png


No signs of EQ errors as well:
1672439251025.png


Clock stability after drift correction also is excellent:
1672439355048.png


@pma, Something is really broken in left channels as it is many orders of magnitude worse than right channels in all aspects.
 
OP
pkane

pkane

Master Contributor
Forum Donor
Joined
Aug 18, 2017
Messages
5,626
Likes
10,202
Location
North-East
@pkane, I'm still using DW 2.0.2 and I'm getting completely different results (without nonlin correction) than what you showed.... and now I found why:

I've used right channels for analysis (probably as selected from the last test I did) and this behaves much better than the left to left channel comparison where I now get similar results than you and those indicate somethning is broken in the DAC-->ADC chain for the left channels which then also explains the ABX.

Right channel results (nonlin cor. off):
View attachment 253563
There is passband ripple of the ADC and DAC but otherwise a close match.

And it is a linear phase ripple because phase is clean:
View attachment 253564

No signs of EQ errors as well:View attachment 253565

Clock stability after drift correction also is excellent:
View attachment 253566

@pma, Something is really broken in left channels as it is many orders of magnitude worse than right channels in all aspects.
That sounds like the E1DA ADC was set to mono mode, where it produces a different result in left and right channels. The left channel is a combination of left and right to reduce noise, and the right channel is just the right channel in that case.
 

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
Listening to files then shows what went wrong here: The original file has that 60'ies ping-pong stereo image whereas the recording has narrowed image on one side, the right side, the guitar isn't fully right. This means the left channel got corrupted by being the mix of left and right channels, that is, the mono mix. Bingo. Verified by doing the same in Audition, now same sound than the recording.

This also explains the "screwed up" result in DW perfectly.

@pma, so not a broken hardware, just a setting error.
 
Last edited:

Grooved

Addicted to Fun and Learning
Joined
Feb 26, 2021
Messages
679
Likes
441
Hello Paul, I have come to an interesting problem today. Normally I prepare my files for listening tests through the soundcard which has the same clock for its ADC and DAC and when comparing the loop product to the original music file, Deltawave gives negligibly low error numbers and the files are indistinguishable in the ABX test. Today, I have decided the loop constituted of Topping D10s DAC (connected to PC through USB-ISO to cut the ground loop) and E1DA Cosmos ADC. For THD or IMD measurements, this loop gives excellent numbers. However, in a loop with a music file, and the result compared to the original file (both previously level matched and time aligned in AA), the result is horrible. Not only in pkmetrics and error distribution, but the difference between the original and the loop product is easily audible. Would you have an idea what could be the issue? It is definitely not the Deltawave itself.
Hi, this is exactly what I've reported some months ago using Tone2 Pro and Cosmos ADC, except that my results were not as bad as yours, so there's certainly a problem somewhere.
I've compared it with my oldest interface (ADC/DAC) only, but I think it would be the same with the newer.
The interface, even if old, has great results with Deltawave, but not in Multitone (THD, Noise,...). Tone2 Pro + ADC are far better in Mutitone, but Deltawave results are lower.

Now, like @KSTR said in older posts, the High Pass can really impact results.
I also tested with different filters on the Tone2 Pro, and I get different results in Deltawave (think it was between 46 and 54dB difference, without correction settings enabled)

One thing I discovered: using the old ADC/DAC interface, I get the same difference numbers with both balanced and unbalanced cables, but the PKMetric is far better with the balanced cables.
Since your D10s is unbalanced, it can lower your PKMetric results too
The lack of syncing means is what kept me from using a Cosmos ADC so far
It was indeed frustrating, and I tried to add SPDIF, TOSLINK, AES, Word clock to get versatility and sync, but without success unfortunately (with the MCLK coming from the ESS chip).
The best I got was with TOSLINK and the interface synced on the optical signal coming from the extension board wired to the Cosmos ADC, but there was random distortion I heard immediately and confirmed once doing a test in Multitone.

EDIT: just your last post @KSTR (I had certainly opened this page before your post but written later), good you found the problem.
Now, how is it possible that the right channel of the recording was like the original right channel, but that the left recorded channel was like the mono mix of both left+right original channels? If it's what you found, what incorrect settings can lead to this? I will listen to the files to understand
 
Last edited:

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
Now, how is it possible that the right channel of the recording was like the original right channel, but that the left recorded channel was like the mono mix of both left+right original channels? If it's what you found, what incorrect settings can lead to this? I will listen to the files to understand
The Cosmos ADC has a mono mode (for 3dB increased signal-to-noise ratio) which works like this:
- Left output (digital) is the mix of left and right analog inputs.
- Right output (digital) is right analog input.
@IVX, To avoid such pitfalls, the right output channel should either be silent or a duplicate of the left output channel or even better yet, the difference/2 of L and R channels... which would then be a L/R to M/S (and vice versa) encoder/decoder.
 
  • Like
Reactions: pma

Grooved

Addicted to Fun and Learning
Joined
Feb 26, 2021
Messages
679
Likes
441
The Cosmos ADC has a mono mode (for 3dB increased signal-to-noise ratio) which works like this:
- Left output (digital) is the mix of left and right analog inputs.
- Right output (digital) is right analog input.
@IVX, To avoid such pitfalls, the right output channel should either be silent or a duplicate of the left output channel or even better yet, the difference/2 of L and R channels... which would then be a L/R to M/S (and vice versa) encoder/decoder.
Ohhh, I certainly never tried to send both L and R from DAC at the same time when using the Mono mode of the Cosmos ADC, I didn't know it was acting like that.
But it's strange that I was sure to get the best results (with the 3dB increase) in Mono mode but using the right channel of the ADC only.

I tested some month ago the Mono mode by sending the DAC right channel only (always the best one) to the Cosmos ADC set in Mono mode:
- to Left only
- to Right only
- to Left+Right with Y XLR cable
I was sure the best results came from feeding the Right channel of the ADC only... I will need to go back and check all the measurements I made
 

syn08

Senior Member
Joined
Aug 9, 2019
Messages
376
Likes
461
Location
Toronto, Canada
To avoid such pitfalls, the right output channel should either be silent or a duplicate of the left output channel or even better yet, the difference/2 of L and R channels... which would then be a L/R to M/S (and vice versa) encoder/decoder.
According to my ES9822 experience, such configurations are not supported. There isn’t even a direct way to chain ES9822 ADCs, so that you would get a dual mono configuration (I did it with an external mux). The only chaining config available is for TDM.
 

IVX

Major Contributor
Joined
Mar 3, 2019
Messages
1,407
Likes
2,771
Location
South of China, SHZ area, - Слава Україні
The Cosmos ADC has a mono mode (for 3dB increased signal-to-noise ratio) which works like this:
- Left output (digital) is the mix of left and right analog inputs.
- Right output (digital) is right analog input.
@IVX, To avoid such pitfalls, the right output channel should either be silent or a duplicate of the left output channel or even better yet, the difference/2 of L and R channels... which would then be a L/R to M/S (and vice versa) encoder/decoder.
this is not my idea, ESS made that this way.
 

pma

Major Contributor
Joined
Feb 23, 2019
Messages
4,591
Likes
10,727
Location
Prague
@KSTR

Hello Klaus, thank you. It is exactly as you have stated.
L = (L+R)/2
R = R
Mystery resolved.

Below are 2-channel measurements with L=100%/R=0% and L=0%/R=100% sine signals. Sometimes it is good to measure everything after listening :):facepalm:.

L onlyL_only.png R onlyR_only.png

After setting Cosmos in the Control Panel to "stereo" (slider 40%) the issue has gone.

E1DA_stereo.png
 
Last edited:

KSTR

Major Contributor
Joined
Sep 6, 2018
Messages
2,690
Likes
6,012
Location
Berlin, Germany
According to my ES9822 experience, such configurations are not supported.
Understood. I had assumed this mixing is done in software, after the conversion, in the XMOS USB interface custom(?) firmware.
 
Top Bottom