DSD vs PCM resolution

[Tony Lauck]

I wonder what happens with an impulse response during these sample rate conversions...

 

You would need to band limit the impulse response to well below 22050 Hz for resampling to work. An actual impulse has infinite bandwidth, or if it is a 1 sample pulse in 44100 format it has an equal energy distribution all the way from 0 upto (but not including) 22050 Hz. If you attempt to resample with lots of energy close to 22050 then there will be imaging and aliasing problems with any practical filter. That is the reason why I apodized the original format.

 

If you used the same procedure that I used but started with a one sample pulse in stead of music, the apodizing process would have given you a waveform that was band limited to 18 kHz and thus safely convertible to/from the 44 kHz format. It could be any 44/24 file you like, so long as a little headroom is left to allow for any possible intersample peaks, sine waves, music, single sample pulses, etc... (That's because the file is the sum of a sequence of these pulses and the processing is linear. True with 24 bit PCM being done in the computer where intermediate calculations are done to much higher accuracy and then the results put back into 24 bit format.)

[audiventory]

After downsampling any higher sample rate to 32 kHz original range 16-18 kHz will mirrored and shifted to 14 kHz. Now it will in band 14-16 kHz.

 

It is fairly for any filter.

 

After such downsampling impossibly restore 16-18 kHz range. It will as distortions in 14-16 kHz range.

[Maldur]

Downsampling to a 32 kHz sampling rate was a convenient way to cut off audio above 16 kHz and the filter that I used was not terribly steep, hence there was still detectable energy up to 18 kHz.

 

(16 kHz is the Nyquist frequency for a sample rate of 32 kHz.)

 

So, there is life after Nyquist frequency? This something new to me, with all respect... but I not buy this.

[Tony Lauck]

So, there is life after Nyquist frequency? This something new to me, with all respect... but I not buy this.

 

Perhaps my description of the process to generate the test file was confusing. Rather than try to justify what I had done, I redid the generation of the test file a different way so as to remove any confusion. I took the base 88/24 track and downsampled it directly to 44/24 using a large offset of the filter so as to strip off all frequencies above 18 kHz, producing a similar spectrum as the original test file, but this time without creating any aliases of images. The result was a normal 44/24 version of the hi-res original, except that it had been apodized (i.e. energy in the range 18-22 kHz removed).

 

Converting the new test file from 44/24 to 192/24 and 44/24 and comparing gave the same null results. See the previous posts for the numbers. The summary is that the difference created by upsampling and downsampling was equivalent to 24 bit dither noise. This is not rocket science, it is just good implementation of digital filters using accurate computer code and an application of the Sampling Theorem.

[Miska]

Why is this relevant to DSD? It's because I do not believe it is possible to pass similar tests using DSD64 and probably also DSD128.

 

Have you tried it? Belief doesn't matter.

 

It's also relevant because it shows that a 23 bit null shows up in spectral plots at -175 dB level, basically showing that comparisons at -130 dB indicate junk performance.

 

Are you referring to my plots? At that point you were specifically questioning DSD capability of reproducing 44.1/16 RedBook resolution and my plots were all from TPDF-dithered 16-bit source converted to 16-bit destination end result. That was what you were talking about at the time...

 

Here's some more fun.

 

Source is 44.1/32 TPDF dithered complex test signal I use, with mix of different amplitude and frequency modulated sines mixed with dirac delta pulses (1 kHz frequency modulated, 7.5 kHz amplitude modulated and 11.5 kHz frequency&amplitude modulated).

 

Conversion of that source to 192/24 (TPDF) and then to 44.1/24 (TPDF):

 

Conversion of the same source to DSD256 and then to 705.6/32 (TPDF) and then to 44.1/32 (TPDF):

 

You can see that DSD256 performance far exceeds 24-bit PCM resolution in audio band.

[Jay-dub]

Why is this relevant to DSD? It's because I do not believe it is possible to pass similar tests using DSD64 and probably also DSD128. It's also relevant because it shows that a 23 bit null shows up in spectral plots at -175 dB level, basically showing that comparisons at -130 dB indicate junk performance.

 

In the previous post it was a 15 bit null that you were claiming is impossible with DSD as the intermediate format.

[Tony Lauck]

In the previous post it was a 15 bit null that you were claiming is impossible with DSD as the intermediate format.

 

Difference in metrics. RMS vs. worst-case.

[Miska]

This is a pretty good practical demonstration of the Sampling Theorem at work using commercially available software on real music. I deliberately chose the two sampling rates as non-integer multiples because some people have questioned the possibility of upsampling and downsampling with non-integer formats.

 

I suspect it would be possible to get even deeper nulls by using 32 bit integer format and I don't believe it would be necessary to use such brutal apodizing of the original signal if I had used a sharper filter. If anyone is curious about other possibilities I suggest they do similar experiments.

 

Now what you need to do is to take a 44.1/24 source file of single sample dirac delta pulses, convert it back and forth (to 192/24 and back) and run the null test against the original... You get to test the actual sampling theorem. Also listen to the converted files made with different converters. To make it more audible, use 22.05 kHz intermediate sampling rate and convert back to 44.1 kHz for listening.

 

Generally you want to diff and listen to the part removed in the your brutal apodizing process.

 

None of this brutal process is needed when your source file/recording is in DSD format and you just play it back as DSD too.

 

For more interesting test, try to run the NULL test through DAC/ADC loop and see which format performs best.

[Tony Lauck]

Have you tried it? Belief doesn't matter.

 

 

 

Are you referring to my plots? At that point you were specifically questioning DSD capability of reproducing 44.1/16 RedBook resolution and my plots were all from TPDF-dithered 16-bit source converted to 16-bit destination end result. That was what you were talking about at the time...

 

Here's some more fun.

 

Source is 44.1/32 TPDF dithered complex test signal I use, with mix of different amplitude and frequency modulated sines mixed with dirac delta pulses (1 kHz frequency modulated, 7.5 kHz amplitude modulated and 11.5 kHz frequency&amplitude modulated).

 

Conversion of that source to 192/24 (TPDF) and then to 44.1/24 (TPDF):

[ATTACH=CONFIG]20283[/ATTACH]

 

Conversion of the same source to DSD256 and then to 705.6/32 (TPDF) and then to 44.1/32 (TPDF):

[ATTACH=CONFIG]20284[/ATTACH]

 

You can see that DSD256 performance far exceeds 24-bit PCM resolution in audio band.

 

I see nice input and output waveforms. But this can be the resolution of the graph. Perhaps you could difference the two files and post the differences. Or better, post the two 44/32 wav files so we could see the actual errors created by the conversions (for each of your separate conversion tests).

[Miska]

Difference in metrics. RMS vs. worst-case.

 

What does that have to do with 16 vs 24 -bit?

 

Anyway it is always nice to play games in pure digital domain, but when you run the stuff through real world converters. Send the data to some real world "PCM" converter chip that has digital filters with at most -120 dB stop-band attenuation (Sabre and few others) and constrained modulators and things don't look as rosy anymore.

 

While one of my filters perform like this:

[ATTACH=CONFIG]20285[/ATTACH]

[Miska]

I see nice input and output waveforms. But this can be the resolution of the graph.

 

Can you explain how it can be resolution of the graph? Both are obtained using same parameters, from a file with same sampling rate.

 

Certainly there are differences at least with the single sample dirac pulses due to different digital filters. You'll get the same with iZotope too if you change any parameters and compare conversions of the pulses.

 

Here, try conversion with this and diff against the original. You can also try to diff between to/from 176.4/192 or with different parameters.

http://www.sonarnerd.net/tmp/dpulse.zip

[Maldur]

While one of my filters perform like this:

[ATTACH=CONFIG]20285[/ATTACH]

 

Miska, this attachment is somehow invalid, gives me error "Message:

Invalid Attachment specified. If you followed a valid link, please notify the administrator".

[mansr]

Here's some more fun.

 

Source is 44.1/32 TPDF dithered complex test signal I use, with mix of different amplitude and frequency modulated sines mixed with dirac delta pulses (1 kHz frequency modulated, 7.5 kHz amplitude modulated and 11.5 kHz frequency&litude modulated).

 

Conversion of that source to 192/24 (TPDF) and then to 44.1/24 (TPDF):

[ATTACH=CONFIG]20283[/ATTACH]

 

Conversion of the same source to DSD256 and then to 705.6/32 (TPDF) and then to 44.1/32 (TPDF):

[ATTACH=CONFIG]20284[/ATTACH]

 

You can see that DSD256 performance far exceeds 24-bit PCM resolution in audio band.

 

Any chance you could share that test file?

[Miska]

Miska, this attachment is somehow invalid, gives me error "Message:

Invalid Attachment specified. If you followed a valid link, please notify the administrator".

 

OK, it still shows up for me, but doesn't embed anymore...

 

In case it is still acting up, I uploaded it here too:

http://www.sonarnerd.net/tmp/filter-resp-31.png

[Maldur]

For me this works now well, kiitos Miska!

[mansr]

OK, it still shows up for me, but doesn't embed anymore...

[ATTACH=CONFIG]20288[/ATTACH]

 

In case it is still acting up, I uploaded it here too:

http://www.sonarnerd.net/tmp/filter-resp-31.png

 

May we see a close-up of the passband?

[Hiro]

You would need to band limit the impulse response to well below 22050 Hz for resampling to work. An actual impulse has infinite bandwidth, or if it is a 1 sample pulse in 44100 format it has an equal energy distribution all the way from 0 upto (but not including) 22050 Hz. If you attempt to resample with lots of energy close to 22050 then there will be imaging and aliasing problems with any practical filter. That is the reason why I apodized the original format.

 

Just like I thought...

[Hiro]

Here's some more fun.

 

Source is 44.1/32 TPDF dithered complex test signal I use, with mix of different amplitude and frequency modulated sines mixed with dirac delta pulses (1 kHz frequency modulated, 7.5 kHz amplitude modulated and 11.5 kHz frequency&amplitude modulated).

 

Conversion of that source to 192/24 (TPDF) and then to 44.1/24 (TPDF):

[ATTACH=CONFIG]20283[/ATTACH]

 

Conversion of the same source to DSD256 and then to 705.6/32 (TPDF) and then to 44.1/32 (TPDF):

[ATTACH=CONFIG]20284[/ATTACH]

 

You can see that DSD256 performance far exceeds 24-bit PCM resolution in audio band.

 

Thanks for posting these measurements. Though, we have to remember that 32bit flash/ladder PCM converters are impossible to make due to implementation issues, and even the 24/20-bit ones often suffer from nonlinear distortions at ~16 bit levels.

[Miska]

May we see a close-up of the passband?

 

You mean ripple?

 

 

Or here, in case there's a problem with the uploaded image:

http://www.sonarnerd.net/tmp/filter-ripple-31.png

 

Y-scale is in dB in this case too. So the pass-band ripple is <0.000000005 dB.

[audiventory]

Yes. It's right filter features. Good resampling filter must have close features:

 

Stop band no worse -200 dB and so low passband ripples.

 

Thus, with using same filter, resampling is not most loss processing in music producction and adjusting audio file resolution to DAC.

 

However, filter has complex of linked features: length, ringing, speed processing,... For minimal phase filter also need minimize phase distortions too.

 

Main secret of filters - art of development - connect these ambigous parameters better way.

[ilok]

In practice, R2R NOS playing PCM sounds infinitely better than DSD native DAC playing DSD.

 

Stop worrying and just enjoy.

[Maldur]

I think, PCM format itself has limitations and weakness, no matter how and what DAC plays PCM - especially when it plays redbook as R2R NOS usually does.

[Paul R]

In practice, R2R NOS playing PCM sounds infinitely better than DSD native DAC playing DSD.

 

Stop worrying and just enjoy.

 

Or just listen to an enjoy that great DSD DAC playing DSD music that sounds like music, not resistors.

 

-Paul

[audiventory]

I think, PCM format itself has limitations and weakness, no matter how and what DAC plays PCM - especially when it plays redbook as R2R NOS usually does.

 

 

Format DSD = PCM + Noise

 

 

Advantages PCM under DSD (as format) is:

 

1. More fast processing (due less sample rate).

 

2. Instant ability to non-linear processing in music production (due absent of the noise).

 

 

 

Advantages DSD under PCM is:

 

1. Simpler transmit via serial communication channels / intrfaces.

 

2. Simpler DAC for same quality.

 

 

 

 

 

Most weakness and high cost of "native" PCM DAC in number resistors that need:

 

1. Right pick up resistors and either amplifier, switcher chips or transistors or other active components,

 

2. Provide work that all in system, assembled on board, with power supply unit, with output device (pre-amplifier),

 

3. Provide temperature stability for all.

 

If assembled 1 DAC only, it is problem.

 

When assembled N*100 it is enough big problem.

 

For N*1000000 DACs it is great problem.

 

 

 

For DSD DAC need provide 2 levels only.

[Maldur]

Yuri, in simple words - most weakness of "native" PCM DAC is hidden in the fact: there is no real or "native" PCM signal whatsoever - all PCM content is downsampled and interpolated from SDM ADC first...

After that fact, these problems with resistors or temperature stability is actually like sort of sport .