Beta-test: DeltaWave Null Comparison software

[pkane]

Testing 1.05
I have problems comparing dsd 64 vs flac 24/192 (1 min, same track, no dither), residual drift error too large.
Continue anyway shows a big step on matched spectra, if the reference is DSF the step is on audio range. But if FLAC is reference the step is on ultrasonic range.
The spectra matches well on audio but it will never over 30 kHz due to DSD nature...
WASAPI or ASIO only playing at 44.1

DSD64 is something I’ve not tried, all my testing was with DSD256. I’ll take a look, but may need your help to upload the two files to share with me, if I can’t reproduce the problem.

It sounds like the files are being downsampled to 44.1KHz. Is there any chance the FLAC file is not at 192KHz? Deltawave will downsample to the lowest rate of the two files, unless you set the option to upsample.

Regards,

-Paul

 

[pkane]

DSD64 is something I’ve not tried, all my testing was with DSD256. I’ll take a look, but may need your help to upload the two files to share with me, if I can’t reproduce the problem.

By the way, here is DSD128 recording (blue) compared to PCM 96/24. Notice the 0 drift and almost 109dBA correlated null. What's more the files match bit-for-bit for 68% of the samples when reduced to 16 bits, so these are really close!

 

[pkane]

Testing 1.05
I have problems comparing dsd 64 vs flac 24/192 (1 min, same track, no dither), residual drift error too large.
Continue anyway shows a big step on matched spectra, if the reference is DSF the step is on audio range. But if FLAC is reference the step is on ultrasonic range.
The spectra matches well on audio but it will never over 30 kHz due to DSD nature...
WASAPI or ASIO only playing at 44.1

Tried to repeat the test with DSD64 and it doesn't look nearly as perfect. Either my demodulator or that format are not matching PCM as well. Perhaps it's just too low a sampling frequency. Or maybe I need to look into using a simple low-pass filter for DSD decoding, as others have suggested. Here's the comparison of the spectra of the two files (red is DSD64, blue PCM 192/24):

Low-pass filtered to 24KHz, here's the matching result for the same two files:

 

[Rock Rabbit]

Tried to repeat the test with DSD64 and it doesn't look nearly as perfect. Either my demodulator or that format are not matching PCM as well. Perhaps it's just too low a sampling frequency. Or maybe I need to look into using a simple low-pass filter for DSD decoding, as others have suggested. Here's the comparison of the spectra of the two files (red is DSD64, blue PCM 192/24):

View attachment 22657


Low-pass filtered to 24KHz, here's the matching result for the same two files:
View attachment 22658

That's the problem, the algorithm needs to correlate only at low frequencies (better precision at very low freq)
Here are my test files https://mega.nz/#F!M5U2xAgQ!TiWHL3GoXbOtQGPfmhpPjg

 

[pkane]

That's the problem, the algorithm needs to correlate only at low frequencies (better precision at very low freq)

Actually, DW does that if you turn on low-pass filtering as shown in the plot I posted. I don't want to make the decision as to what matches better ahead of time, but the choice is yours. You get to decide if you want to apply a low pass (or a high pass) filter, and where to place it. By default, DW attempts to match the whole spectrum of the file.

Thanks for uploading the files, I'll take a look and report back.

 

[Rock Rabbit]

The calculated spectrums seems to show lower amplitude than normal
With a single FLAC 24/192 the correlated null is almost 293 dB, but the spectrum of delta shows progressive error below 20 kHz

 

[pkane]

The calculated spectrums seems to show lower amplitude than normal
With a single FLAC 24/192 the correlated null is almost 293 dB, but the spectrum of delta shows progressive error below 20 kHz

Sorry, I'm not sure I understand. Can you please post a screen shot of what you mean? Are you comparing one file to itself, or...?

 

[Blumlein 88]

Actually, DW does that if you turn on low-pass filtering as shown in the plot I posted. I don't want to make the decision as to what matches better ahead of time, but the choice is yours. You get to decide if you want to apply a low pass (or a high pass) filter, and where to place it. By default, DW attempts to match the whole spectrum of the file.

Thanks for uploading the files, I'll take a look and report back.

That brings me to a feature request. Sometimes being able to do a low pass and high pass together would be desirable.

I've played with doing one or the other in software on the files prior to comparing, but having it right there as a choice would be much more convenient. Pretty please?

 

[pkane]

That brings me to a feature request. Sometimes being able to do a low pass and high pass together would be desirable.

I've played with doing one or the other in software on the files prior to comparing, but having it right there as a choice would be much more convenient. Pretty please?

Funny, I was thinking that this would be useful earlier today

 

[pkane]

Testing 1.05
I have problems comparing dsd 64 vs flac 24/192 (1 min, same track, no dither), residual drift error too large.
Continue anyway shows a big step on matched spectra, if the reference is DSF the step is on audio range. But if FLAC is reference the step is on ultrasonic range.

The large error indicates a matching problem. Continuing after this warning, you are unlikely to get good results since the phase matching algorithm didn't converge. The clock drift error reported exceeded 500ppm, but the actual error was much larger! The worst part was that it's not a constant clock drift. If you take a look at the DSD64 comparison I posted here, the clock drift was 0.01ppm. It would appear that your files are much harder to match due to the large, variable clock difference. Do you know how the DSF file was generated?

The timing offset error between successive chunks of the file is as follows (in whole samples!) There's nothing DW can do to correct for this kind of a clock drift, unfortunately...

Correlation error between 20 consecutive parts of the two files:

 

[Rock Rabbit]

Sorry, I'm not sure I understand. Can you please post a screen shot of what you mean? Are you comparing one file to itself, or...?

Using only the FLAC test file, the low freq spectrum is below -50 dB, in another program (Izotope Rx) below 400 Hz it´s over -50 dB (same FFT size, same window).

Maybe I could suggest log scale and full freq range in one page (when not zoomed) for spectrum.
And many thanks for the program and analysis , the test waves were downloads from a Hi-Res site

 

[pkane]

Using only the FLAC test file, the low freq spectrum is below -50 dB, in another program (Izotope Rx) below 400 Hz it´s over -50 dB (same FFT size, same window).

Maybe I could suggest log scale and full freq range in one page (when not zoomed) for spectrum.
And many thanks for the program and analysis , the test waves were downloads from a Hi-Res site

That's likely a bug I just fixed in v1.0.6 -- please try the update.

 

[pkane]

For your testing pleasure, DeltaWave version 1.0.6b is now available.

Some bug fixes, enhancements, and new features:

• Improved plot axis labeling with a more intelligent choice of units and decimal places to display
• Added AB/X style Comparator for sighted and blind auditioning of the files being compared
• Added the support for DFF-formated DSD files (uncompressed only)
• Added playback position display and ability to set playback position by clicking on the control
• Added the option to turn off MD5 hash in the results, as this could cause some delay when processing large files
• Changed how the Trim functions work: the data that is trimmed is now skipped when reading the file. Previously the whole file was loaded into memory and then trimmed
• Added a 50% bit-perfect match result. This is how many bits the files need to be reduced to in order for 50% of the samples to match exactly
• Added support for two simultaneous filters. A low-pass and a high-pass filter can now be applied at the same time

As always, your comments, suggestions, and bug reports are more than welcome.

Regards,

-Paul

 

[pkane]

That's the problem, the algorithm needs to correlate only at low frequencies (better precision at very low freq)
Here are my test files https://mega.nz/#F!M5U2xAgQ!TiWHL3GoXbOtQGPfmhpPjg

Just used a freebie PCM to DSD encoder to convert your FLAC file to DSD64 (don't know anything about the quality of their modulator or filters). The result appears to be a whole lot better than the DSD file you downloaded from the hi-res website, so I have to assume something's wrong with that file:

 

[Rock Rabbit]

Testing 1.06
Two FLACs 24/44.1,... in the original or matched spectra windows the spectrum doesn´t goes down at Nyquist (the waves are band limited ok)
Now when playing start the green progress bar everytime,... and hang the reproduction when using ASIO

 

[pkane]

Testing 1.06
Two FLACs 24/44.1,... in the original or matched spectra windows the spectrum doesn´t goes down at Nyquist (the waves are band limited ok)
Now when playing start the green progress bar everytime,... and hang the reproduction when using ASIO

Can you please post a screen shot of the spectrum not going down?

 

[Rock Rabbit]

 

[pkane]

View attachment 22708

So why would you expect that this would go down below about -120dB level, where it appears to be? Band limited doesn't mean that the values must go down before the Nyquist frequency, that only happens if the signal naturally drops off or is low-pass filtered. Looks like -120dB is the noise floor of the signal, and so it doesn't drop any further.

 

[Rock Rabbit]

Oops, the signal goes down so slow, my apologies
I have serious issues with the colors and background, but I don´t read correctly the scale of freq

 

[pkane]

Oops, the signal goes down so slow, my apologies
I have serious issues with the colors and background, but I don´t read correctly the scale of freq

I’m curious, what color combinations might work better? I didn’t think about picking a high contrast set, but sounds like that might help.