How can a non-oversampling DAC sound good?

[Fokus]

1) Have a look at the sampling theorem. The ideal reconstruction filter is Sinc(). Using this guarantees that the pre-sampling band-limited signal is fully reconstructed. If the original signal had no ringing, and was band-limited, then the post-DAC signal will also have no ringing, and this despite Sinc() ringing from the Big Bang to Armageddon. That should give a clue.

 

2) Most recording ADCs and mastering sample rate convertors are half-band linear phase filters. These ring. It is this ringing that makes it through, both in the case of a Sinc-filtered DAC and in the case of a NOS DAC. Create an impulse in a DAW at a high sample rate. Downsample it with linear phase half-band filter. Play it through the NOS DAC. Observe.

 

If the ADC/mastering filter is half-band (most of them are, though it is getting better) then it violates the sampling theorem by allowing content smack up to Fs/2 (in actuality it aliases during recording, even worse). This content is illegal and will cause a Sinc() DAC to ring itself. But again: due to an illegal input. The filter itself is not wrong.

 

Recipe for success: make recordings with sufficient attenuation at Fs/2 (like -60dB or more), using anti-alias filters with a transition band of some 4kHz wide. All problems solved.

[eintom]

Here is a great paper by Metrum on NOS DACs, that explains why they measure badly:

http://www.metrum-acoustics.com/Design%20Philosophy%20Metrum%20Acoustics.pdf"

 

i use such a Metrum Acoustics NOS DAC and i love it. :-)

before i had an Auralic Vega, which was not bad too. But the natural sounding Metrum beats it in any aspect, it sound more realistic to me and i do not care about measurements ;-)

[manisandher]

There should be no new aliasing effects below 22.05 kHz, as the Prism's anti-alias filter will have eliminated this. However, the Altmann's imaging effects around 44.1kHz will still be there, but totally cut off above 22.05 Khz. Below this, all the imaging effects should remain, as in the original plot.

Well, it seems it isn't quite as simple as that.

 

Here's the Altmann's output of a 15 kHz tone:

 

 

Here's what I thought I'd get with the output going through the Prism ADC set to 16/44.1:

 

 

But in actuality it looks like this:

 

 

It seems that the Prism's anti-alias filter has indeed eliminated the effects of the Altmann's imaging below 20.05 kHz. I don't understand how this could be the case.

 

Mani.

[manisandher]

1) Have a look at the sampling theorem. The ideal reconstruction filter is Sinc(). Using this guarantees that the pre-sampling band-limited signal is fully reconstructed. If the original signal had no ringing, and was band-limited, then the post-DAC signal will also have no ringing, and this despite Sinc() ringing from the Big Bang to Armageddon. That should give a clue.

 

2) Most recording ADCs and mastering sample rate convertors are half-band linear phase filters. These ring. It is this ringing that makes it through, both in the case of a Sinc-filtered DAC and in the case of a NOS DAC. Create an impulse in a DAW at a high sample rate. Downsample it with linear phase half-band filter. Play it through the NOS DAC. Observe.

 

If the ADC/mastering filter is half-band (most of them are, though it is getting better) then it violates the sampling theorem by allowing content smack up to Fs/2 (in actuality it aliases during recording, even worse). This content is illegal and will cause a Sinc() DAC to ring itself. But again: due to an illegal input. The filter itself is not wrong.

 

Recipe for success: make recordings with sufficient attenuation at Fs/2 (like -60dB or more), using anti-alias filters with a transition band of some 4kHz wide. All problems solved.

Thanks Fokus. Let me digest all this.

 

Mani.

[manisandher]

i use such a Metrum Acoustics NOS DAC and i love it. :-)

before i had an Auralic Vega, which was not bad too. But the natural sounding Metrum beats it in any aspect, it sound more realistic to me and i do not care about measurements ;-)

 

I've never heard a Metrum, but have read lots of good reports. And I quite like the sound of the Altmann too. I just couldn't square how it's captured output could sound more accurate (as opposed to more pleasing) than that of a 'properly-engineered' DAC such as the 2Qute. And hence why I started this thread.

 

Mani.

[manisandher]

If the original signal had no ringing, and was band-limited, then the post-DAC signal will also have no ringing...

OK, so what's happening with measurements such as those on Stereophile? Below are some measurements from the Mytek Brooklyn review...

 

Typical FIR reconstruction filter:

 

And resulting impulse response:

 

 

So called 'Slow Roll-off' filter:

 

 

And resulting impulse response:

 

 

I'm assuming there was no ringing in the impulse that the Audio Precision created. The resulting ringing in the two impulse responses is due entirely to the type of reconstruction filter used.

 

And I assumed that if it one were to eliminate the filter altogether it would result in a terrible spectrum (loads of imaging), but a perfect impulse response (with zero ringing). Is this wrong?

 

Mani.

[Fokus]

I'm assuming there was no ringing in the impulse that the Audio Precision created.

 

True. But that impulse is also not band-limited, thus a violation.

[audiventory]

Yep, happy to do this - I'm interested in knowing too. The biggest issue will be that the plots will only extend to 22.05 kHz and not 96kHz. For the Altmann, I would expect exactly this (plot stopping at 22.05 kHz):

 

[ATTACH=CONFIG]32838[/ATTACH]

 

There should be no new aliasing effects below 22.05 kHz, as the Prism's anti-alias filter will have eliminated this. However, the Altmann's imaging effects around 44.1kHz will still be there, but totally cut off above 22.05 Khz. Below this, all the imaging effects should remain, as in the original plot.

 

Me seems better way is using of pure sine, generated aftifically in a sound editor in 24 bit resolution. Because in the test signal too many components that make analysis difficult.

[audiventory]

1) Have a look at the sampling theorem. The ideal reconstruction filter is Sinc(). Using this guarantees that the pre-sampling band-limited signal is fully reconstructed.

 

Band limited and time unlimited.

[manisandher]

Me seems better way is using of pure sine, generated aftifically in a sound editor in 24 bit resolution. Because in the test signal too many components that make analysis difficult.

Yes, the test signal does seem a bit noisy. I might look into redoing things with a generated test signal at a later date.

 

Mani.

[manisandher]

Band limited and time unlimited.

Yes, and it's impossible to have both simultaneously. So 'the perfect' filter simply doesn't exist.

 

I've always found HQPlayer fascinating to play around with. Each filter really does seem to have its own sonic character and it's sometimes difficult deciding which one to choose. I've also found the Altmann quite educational - it sounds better than I thought it would.

 

Mani.

[audiventory]

Yes, and it's impossible to have both simultaneously. So 'the perfect' filter simply doesn't exist.

 

I've always found HQPlayer fascinating to play around with. Each filter really does seem to have its own sonic character and it's sometimes difficult deciding which one to choose. I've also found the Altmann quite educational - it sounds better than I thought it would.

 

Filter is balanced system of the features:

1. Steepness of transient between pass- and stop-band.

2. Suppressing in stop-band.

3. Ringing level.

4. Delay between input and output (for finite impulse response).

5. Pre-ringing and post-ringing energy. As rule, total ringing energy is constant. We can only push balance from pre- to post-.

6. Dispersion of levels in passband.

[Jud]

The non-filtering NOS DAC I listened to was playing (only) 16/44.1 WAV ripped from CDs. So there would be no ultrasonics (at least above 22.05KHz) to image. Everything coming through would have been high "audible" (usually thought of as up to 20K, but I might hear up to 16K max) or very near ultrasonic. So is the Prism doing that much better a job than whatever ADC was used for these CDs, and that accounts for the whole of the subjective difference ("accurate" vs. "teeth-gritting")?

[manisandher]

The non-filtering NOS DAC I listened to was playing (only) 16/44.1 WAV ripped from CDs. So there would be no ultrasonics (at least above 22.05KHz) to image.

 

Jud, any NOS/filterless DAC playing a CD rip will image like made above 22.05 kHz, even if there are no ultrasonics in the source material.

 

Take a look at this plot again:

 

 

This is a 15 kHz test tone - no signal above 22.05 kHz. Look at the spectrum from 44.1 kHz down to 22.05 kHz. Can you see that it's a perfect mirror image of the spectrum from 0 to 22.05 kHz? And also that 44.1 kHz to 66.15 kHz is a perfect copy of 0 to 22.05 kHz? (Not exactly a perfect copy because the Tascam's anti-alias filter is starting to kick in to cut everything before 96kHz.) That's imaging. The non-filtering NOS DAC you listened to would have been doing exactly the same, so loads of energy above 22.05 kHz. Audible? Well if not directly, almost certainly indirectly.

 

Mani.

[Jud]

Jud, any NOS/filterless DAC playing a CD rip will image like made above 22.05 kHz, even if there are no ultrasonics in the source material.

 

Take a look at this plot again:

 

[ATTACH=CONFIG]32858[/ATTACH]

 

This is a 15 kHz test tone - no signal above 22.05 kHz. Look at the spectrum from 44.1 kHz down to 22.05 kHz. Can you see that it's a perfect mirror image of the spectrum from 0 to 22.05 kHz? And also that 44.1 kHz to 66.15 kHz is a perfect copy of 0 to 22.05 kHz? (Not exactly a perfect copy because the Tascam's anti-alias filter is starting to kick in to cut everything before 96kHz.) That's imaging. The non-filtering NOS DAC you listened to would have been doing exactly the same, so loads of energy above 22.05 kHz. Audible? Well if not directly, almost certainly indirectly.

 

Mani.

 

I was thinking I must be an idiot, but thank you (really) for pointing out precisely how!

 

 

Sent from my iPhone using Computer Audiophile

[manisandher]

Hey Jud, just cadge a lift with our PM as she heads back to the UK and buy me that beer you've always promised :-)

 

Mani.

[YashN]

These distortions, even though ultrasonic, must have some effect on the sound for the reasons you cite. Here's a linear scale plot of the Altmann at 15kHz:

 

[ATTACH=CONFIG]32837[/ATTACH]

 

The imaging around 44.1 kHz is clearly seen.

Also check the graph in Log scale, there could be a lot of energy spent in the lower regions, measurement results for sub-20Hz may be surprising sometimes.

[YashN]

As audiventory points out, a non-oversampling DAC will be sending a whole bunch of ultrasonic grunge to the electronics downstream (its own output stage included).

 

Both are real phenomena.

 

The paper is very interesting in that it shows how what we are used to perform as measurements are imperfect, but so far, the most common ways to perform them.

 

Sometimes, some people lose sight of those imperfections, take assumptions as facts, forget to do correlations, then focus on some narrow measurements.

 

e.g. Focusing on just PSRR and noise readings for a Power Supply.

 

The tendency we have to focus on Frequency Response is another one, when there is a mind-boggling amount of information about sound in the transients.

 

Of course, both FR and the transients are important, as are other things.

[YashN]

It seems that the Prism's anti-alias filter has indeed eliminated the effects of the Altmann's imaging below 20.05 kHz. I don't understand how this could be the case.

Mani, before you measured the DAC captures, did you calibrate the PRISM with REW?

[Jud]

1) Have a look at the sampling theorem. The ideal reconstruction filter is Sinc(). Using this guarantees that the pre-sampling band-limited signal is fully reconstructed. If the original signal had no ringing, and was band-limited, then the post-DAC signal will also have no ringing, and this despite Sinc() ringing from the Big Bang to Armageddon. That should give a clue.

 

2) Most recording ADCs and mastering sample rate convertors are half-band linear phase filters. These ring. It is this ringing that makes it through, both in the case of a Sinc-filtered DAC and in the case of a NOS DAC. Create an impulse in a DAW at a high sample rate. Downsample it with linear phase half-band filter. Play it through the NOS DAC. Observe.

 

If the ADC/mastering filter is half-band (most of them are, though it is getting better) then it violates the sampling theorem by allowing content smack up to Fs/2 (in actuality it aliases during recording, even worse). This content is illegal and will cause a Sinc() DAC to ring itself. But again: due to an illegal input. The filter itself is not wrong.

 

Recipe for success: make recordings with sufficient attenuation at Fs/2 (like -60dB or more), using anti-alias filters with a transition band of some 4kHz wide. All problems solved.

 

Thanks Fokus. Let me digest all this.

 

Mani.

 

 

Don't know why a little sparkling water and a fresh look would make things so much easier to understand, but it seems to've.

 

 

- Band limiting and no ringing in the original mean there's no out-of-band material to make the DAC filter ring and no ringing from the ADC to pass through.

 

- The filtering used in many ADCs rings, and with either no filter or without a DAC filter designed to remove the ringing in the original, that ringing will come through the DAC.

 

- Besides ringing, the filters used in many ADCs don't properly band-limit, so the out-of-band input can make a DAC filter ring.

 

I think that's what Fokus was saying, prior to his recipe for success.

[audiventory]

If DAC have no filters, there half of output power (in linear scale) will wasted.

 

If ADC have no analog and digital filters at input, there will captured infinite frequency range, folded by half input sample rate of the ADC. The sample rate is higher than sample rate at output ADC.

 

What about ringing of a speakers?

 

It is mechanical system. I suppose, that after voltage on/off it will to continue oscillate some time.

[barrows]

In the Altman DAC is there a powerful analog filter in the I/V section? Some NOS DACs use strong analog filtering (even into the audio bandwidth) to get rid of some of the problems.

[jamesg11]

In the Altman DAC is there a powerful analog filter in the I/V section? Some NOS DACs use strong analog filtering (even into the audio bandwidth) to get rid of some of the problems.

Metrum?

[Jud]

In the Altman DAC is there a powerful analog filter in the I/V section? Some NOS DACs use strong analog filtering (even into the audio bandwidth) to get rid of some of the problems.

 

Was thinking about that. Audio Note DACs used a transformer as the filter. There's a pic of the Altmann board on their site if you want to look. But wouldn't some of the posted graphs seem to indicate little if any filtering?

 

 

Sent from my iPhone using Computer Audiophile

[manisandher]

Mani, before you measured the DAC captures, did you calibrate the PRISM with REW?

Yash, I was just being an idiot! Here's the Altmann output again before the ADC:

 

 

And here are the expected and actual outputs after the ADC:

 

 

I thought that the ADC's anti-alias filter had somehow 'corrected' the sin(x)/x drop, and just couldn't understand how this could be because it shouldn't affect anything below 1/2 fs. But of course, it's done no such 'correction'. All we're seeing here is the ADC's 16 bit noise floor! I haven't done it, and have no inclination to do so, but if I were to switch the ADC to its 24 bit setting, I'm certain the expected and actual spectra would look identical.

 

Mani.