"I can hear them church bells pre-ringing!" - A (mostly) time domain investigation into DAC-like reconstruction filters

[modus2000]

you changed my view on how dac works

My next interest - how Macos manage sound inside the system. Is there any way to obtain bit-perfect sound track except Audirvana. Can I receive lossless quality via browser Safari ? As I understood Macos (as well as Win) do re-sampling via 48 khz. How to avoid re-sampling?

Do I really need this bit-perfect?


I would appreciate if somebody explain

 

[voodooless]

My next interest - how Macos manage sound inside the system. Is there any way to obtain bit-perfect sound track except Audirvana. Can I receive lossless quality via browser Safari ? Do I really need this bit-perfect? I would appreciate if somebody explain

Not really something for this topic. Maybe create your own. The most likely answer is no, you will not get bit perfect audio though a browser in MacOS. Do you need it: probably not. The audio subsystem of MacOS had been transparent for years and years. You’ll never hear the difference. And the reason for this does actually touch the very topic discussed here in this thread.

 

[antcollinet]

interesteng, thank you
Previously I thought that modern dac just round the cornenrs ))

If you watch Monty's show and tell I linked above, find the spot where the DAC can perfectly reconstruct a sine wave from just two samples per cycle. Clearly it is not "just rounding the corners.

(Look at 4:30 where he shows 15kHz - just 3 samples per cycle - and a horrible "digital" waveform. But still a perfect sine wave on the ouput. He then ramps it up to 20kHz - just two samples per cycle, still a perfect sine wave.)

The whole video really is an excellent primer for how digital audio works - and you don't need to understand all the maths of the sampling theorem to grasp it.

 

[antcollinet]

Do I really need this bit-perfect?

See my signature

No, you don't.

Correctly implemented manipulations (up sampling, down sampling (to no lower than redbook), volume control) don't have any audible impacts on the sound, unless those changes are the reason for the manipulation (EG equalisation/tone control or volume control etc.)

 

[Mnyb]

As an amateur I see it like this the samples are discrete values and your are fitting a mathematical function true the values the “reconstruction filter” it very similar to,the bandwidth limiting that was done to get the same samples .

There is no stairsteps just discrete values described by a function.

The stairs steps , the “sample and hold” can be intermediate values in some designs , your not supposed to listen to those

 

[Hayabusa]

View attachment 521625
I always find this an excellent illustration.

Also includes the high frequency loss of the sample and hold

 

[voodooless]

If you watch Monty's show and tell I linked above, find the spot where the DAC can perfectly reconstruct a sine wave from just two samples per cycle

To be pedantic: this is not correct. You need more than two samples per cycle. The rule is more than 2x the sample rate, not at least 2x the sample rate That is not just because you need a bit of spectrum for the filter to work. If you have exactly 2 samples per cycle. You can rotate the phase such that both of these would land at the zero value, resulting in no sound at all

 

[danadam]

If you have exactly 2 samples per cycle. You can rotate the phase such that both of these would land at the zero value, resulting in no sound at all

Even if you don't rotate, there is more than one sine of Fs/2 frequency that can be fit into such samples:

Nyquist–Shannon sampling theorem - Wikipedia

en.wikipedia.org

 

[voodooless]

Even if you don't rotate, there is more than one sine of Fs/2 frequency that can be fit into such samples:

Nyquist–Shannon sampling theorem - Wikipedia

en.wikipedia.org

Yeah, sure, but that is much harder to explain

 

[antcollinet]

To be pedantic: this is not correct.

To be even more pedantic - I know. But even then with 44.1kHz and a 20kHz sine - the majority of cycles have only two samples - it is just that they are closer together than the period of the sine. Only around one in 5 cycles will actually have three samples within it. The cycles with only two samples are still perfectly reconstructed (to the limits of quantisation noise)

Or you could say - 2.205 samples per period.

Either way - not just rounding the corners

 

[voodooless]

To be even more pedantic - I know

Of course, you know