Does Phase Distortion/Shift Matter in Audio? (no*)

[bmc0]

The answer to that question is "it's indeterminant" at pi.

Thanks very much for your response. I was thinking that the problem at pi was theoretically solved by the signal being of infinite length. If it isn't, is "ideal" downsampling via the Whittaker-Shannon interpolation formula as explained here mathematically invalid?

which means you can never MAKE that signal from a continuous system that's being sampled

If you start with a continuous-time Dirac delta function, bandlimit it with a normalized sinc function, then sample it at all integer times, do you not then have the signal in question? The central peak is the non-zero sample; all other samples are at the zero crossings.

 

[j_j]

If you start with a continuous-time Dirac delta function, bandlimit it with a normalized sinc function, then sample it at all integer times, do you not then have the signal in question? The central peak is the non-zero sample; all other samples are at the zero crossings.

Strictly speaking, you can do that, but you've still left off the anti-aliasing filter. I think this is getting into pure semantics. What's "signal" and what's "sampling system", and adding that theoretically infinite filter, in truncated form, now are you doing proper sampling? No, you're not, you have out of band signal.

 

[bmc0]

Strictly speaking, you can do that, but you've still left off the anti-aliasing filter.

My understanding was that—in the "pure math" domain only—a sinc filter (not truncated) fulfills the bandlimiting requirement imposed by sampling theorem. I know that this filter is not realizable in practice.

Edit: Actually, disregard the above. I'm fairly convinced now that my thinking about what happens as the filter length goes to infinity was incorrect. There must (I think) still be ambiguity at exactly Fs/2.

 

[amirm]

We should not forget that today's music hardly contains any real and unprocessed ADC capture. Therefore a DAC must cope with signals that an ADC would never produce.

Your processing better not create illegal signals. And certainly not a pure impulse. Pretty sure neither the talent, the engineer or the label would be looking for that.

There are certainly effects that produce aliasing components (and clipping, etc.). If those are desirable artifacts, then sure. Otherwise, folks better know what they are doing in digital domain. You can't expect the system to compensate for these errors.

 

[markstein]

Otherwise, folks better know what they are doing in digital domain. You can't expect the system to compensate for these errors.

This is wishful thinking.

I rather prefer to have a DAC with a certain "immunity" to errors like (intersample) clipping than to hope that all sound engineers know what they are doing...

(They certainly do not, otherwise there wouldn't be so many CDs producing intersample clipping - for no other reason than stupid and brainless loudness maximization)

 

[amirm]

I rather prefer to have a DAC with a certain "immunity" to errors like (intersample) clipping than to hope that all sound engineers know what they are doing...

Then look for them. I would not as it would permanently rob you of dynamic range for the other 99.999% of music that doesn't have that issue.

 

[amirm]

(They certainly do not, otherwise there wouldn't be so many CDs producing intersample clipping)

If you have that much music you care about, simply set a digital volume level to a few dBs lower. Don't ask for a DAC to reduce its dynamic range for everyone else.

 

[Mike123]

I rather prefer to have a DAC with a certain "immunity" to errors like (intersample) clipping than to hope that all sound engineers know what they are doing...

Nobody is talking about depriving DACs of output filters. Filters are already present in all modern DAC chips. This is protection against bad recording, errors in files, and those who like to feed something strange to the device. The point is that these filters serve only for protection. 99.99% of the time, DAC filters do not work because the signal fed to them is correctly formed and does not contain frequencies higher than FS/2.

 

[bmc0]

We should not forget that today's music hardly contains any real and unprocessed ADC capture. Therefore a DAC must cope with signals that an ADC would never produce.
Same thing as with intersample overs. A proper ADC never produces those but that's no sufficient reason why a DAC wouldn't need to care.

Also, many ADC filters have the middle of their transition band at Fs/2. The AKM5578 is one such example that I pointed out previously (see page 11, fig. 3). So much for not violating sampling theorem .

the other 99.999% of music that doesn't have that issue

Intersample overs are pretty common in my experience. As you said, a digital volume control works just fine.

The point is that these filters serve only for protection.

Converting a digital signal to analog without a filter inherently produces a (theoretically infinite) series of images above Fs/2. It does not matter whether the digital signal is "correctly formed" or not—the reconstruction filter is an absolute requirement unless you want heaps of ultrasonic junk in your output signal.

 

[markstein]

If you have that much music you care about, simply set a digital volume level to a few dBs lower. Don't ask for a DAC to reduce its dynamic range for everyone else.

This is exactly what I am doing.

This requires, however, a digital volume control at the correct location in the digital pipeline, especially prior to any SRC. The very popular minidsp SHD, e.g., had the ASRC prior to the digital volume control, which AFAIK in the meantime was changed upon request from many users.

 

[markstein]

My RME ADI-2 Pro e.g. has an option to digitally attenuate the signal prior to SRC by 3dB. This is a good "workaround" as well.

I do not notice this 3dB SINAD reduction at all.

 

[Mike123]

Converting a digital signal to analog without a filter inherently produces a (theoretically infinite) series of images above Fs/2. It does not matter whether the digital signal is "correctly formed" or not—the reconstruction filter is an absolute requirement unless you want heaps of ultrasonic junk in your output signal.

The NOS filter mode exists in some DACs. It does not bring anything terrible except inaudible ultrasonic garbage. No loud unnecessary sounds, oversaturation of analog circuits. Therefore, the filter is, of course, desirable, but not categorically obligatory.

 

[gnarly]

Not really, when one applies what we know about absolute thresholds, physics thresholds, etc, then filters that do not go to zero outside the passband, but that are down 120dB or so, are easily shown to be satisfactory because the error is below the noise floor of the atmosphere at your ear drum.

Thanks.
Does that mean audio discussions, or even measurements, about whatever takes place below 120dB down, .......don't have any chance of significance in terms of audio reality?

 

[NTK]

Also, many ADC filters have the middle of their transition band at Fs/2. The AKM5578 is one such example that I pointed out previously (see page 11, fig. 3). So much for not violating sampling theorem .

It is because the ADC is designed for real life audio applications, which means the upper limit of the usable bandwidth is intended to be 20 kHz, not fs/2 (= 24 kHz). The filter doesn't reach its stop-band spec until 28 kHz. Why? Because in the incoming signal, frequencies > 28 kHz alias back to < 20 kHz. The designers don't care much about what happens between 20 kHz and fs/2.

From: http://www.dspguide.com/ch3/4.htm analog filters for data conversion

 

[pma]

It is because the ADC is designed for real life audio applications, which means the upper limit of the usable bandwidth is intended to be 20 kHz, not fs/2 (= 24 kHz)

Such explanation is difficult to accept. Many acoustical instruments have spectrum far exceeding 20kHz. You guys are trying to find excuses for poor design practices.

 

[NTK]

Such explanation is difficult to accept. Many acoustical instruments have spectrum far exceeding 20kHz. You guys are trying to find excuses for poor design practices.

Then why don't you start your own ADC company and have the whole market to yourself.

 

[kyuu]

Such explanation is difficult to accept. Many acoustical instruments have spectrum far exceeding 20kHz. You guys are trying to find excuses for poor design practices.

Good design should be concerned about the limits of the human auditory system, not about whether some instruments may technically produce frequencies in the ultrasonic band. Unless you're designing a system to play back to your pet bat.

 

[bmc0]

It is because the ADC is designed for real life audio applications, which means the upper limit of the usable bandwidth is intended to be 20 kHz, not fs/2 (= 24 kHz). The filter doesn't reach its stop-band spec until 28 kHz. Why? Because in the incoming signal, frequencies > 28 kHz alias back to < 20 kHz. The designers don't care much about what happens between 20 kHz and fs/2.

Yes, I'm fully aware of why one might choose to design the filter like this and why it's often considered an acceptable tradeoff. However, Amir keeps talking about the importance of strictly obeying the bandlimiting requirement of sampling theorem and seemingly asserting that no real ADC would violate this. Yet here's a top-of-the-line ADC from a respected manufacturer which clearly does (and many other examples can be found).

The filter in question is of the "half-band" type mentioned by @j_j in this comment. A normalized sinc filter is in fact the same thing, just infinitely steep.

Please do correct me if anything I've written in this comment is wrong.

 

[gberchin]

The designers don't care much about what happens between 20 kHz and fs/2.

The wider transition band also benefits the impulse response characteristics of the filter.

 

[danadam]

But is it BW < Fs/2 or BW <= Fs/2 ?

AFAIK it is "<" and there would be ambiguities in case of "<=". The example I read somewhere was that if you happened to sample a signal of Fs/2 frequency at its zero crossing, you would just get a series of zero samples. Wikipedia shows that a series of +1, -1, +1, -1, ... samples would correspond to a whole family of signals of Fs/2 frequency:

Nyquist–Shannon sampling theorem - Wikipedia

en.wikipedia.org