Do upsampling DAC's still need "steep" reconstruction filters?

[Endibol]

For some time I am wondering why modern DAC's are still judged on the steepness of their reconstruction filters. What I understand is that (almost) all DAC-chips nowadays have on-board upsampling functionality. Suppose a common 44.1 kHz signal is upsampled 8x, which would mean a sample rate of 352 kHz. In that case the filter wouldn't need to be so steep as to reduce the signal to approx. say -80 dB at 22 kHz, but only at 176 kHz. This would require a filter with much slower roll-off characteristics. Still I see in all reviews filters with a cut-off at approx. 22 kHz.
Why? Wouldn't a less steep filter not offer all kind of benefits like easier implementation, less phase shift issues etc.?
Who can give my some clarity on this point? Thanks a lot!

 

[NTK]

The reconstruction filter is also the interpolation filter. The idea of over-sampling is to replace the need of a very steep analog anti-imaging filter with a steep digital filter followed by a gentle analog one.

 

[staticV3]

When Amir tests the anti aliasing filter of a DAC at 44.1kHz input, that is always with the DACs internal oversampling turned on.
What you see is actually what you get during real world usage. It's not a synthetic stress test so to say.

 

[musicforcities]

I recommend this video, it is the clearest explanation I have found of how this works mathematically and how the filters work in the process. It also explains, implicitly why ladder DACs measure worse.

 

[kemmler3D]

Questions of aliasing aside, it's good to avoid reproducing ultrasound if you can. A lot of tweeters have nasty breakup modes at 20khz or a bit higher. I can't explain exactly how in physical terms, but distortion at higher frequencies can affect distortion at lower frequencies, so in theory feeding >22khz energy to your tweeter could have audible effects on your treble. This now almost famous paper from Bruno explains / demonstrates it: https://purifi-audio.com/wp-content/uploads/2022/03/220211_R05-Notchfilter.pdf

From this point of view, a steep filter at 22khz is desirable regardless of whether the resampling calls for it.

 

[Endibol]

Questions of aliasing aside, it's good to avoid reproducing ultrasound if you can. A lot of tweeters have nasty breakup modes at 20khz or a bit higher. I can't explain exactly how in physical terms, but distortion at higher frequencies can affect distortion at lower frequencies, so in theory feeding >22khz energy to your tweeter could have audible effects on your treble. This now almost famous paper from Bruno explains / demonstrates it: https://purifi-audio.com/wp-content/uploads/2022/03/220211_R05-Notchfilter.pdf

From this point of view, a steep filter at 22khz is desirable regardless of whether the resampling calls for it.

It is certainly an interesting video. However, my point is that nowadays upsampling is used within the modern DAC's. Therefore, just to avoid imaging the filters don't need to be steep anymore. But you still see steep filters in the designs.. So my question is, why?

 

[solderdude]

It is certainly an interesting video. However, my point is that nowadays upsampling is used within the modern DAC's. Therefore, just to avoid imaging the filters don't need to be steep anymore. But you still see steep filters in the designs.. So my question is, why?

To prevent mirror images in the audible domain.
The upsampling is needed because there are less bits available to work with (usually somewhere between 4 and 6) .
So the filter MUST be steep. Or should I say should be steep.
When one uses software upsampling the upsampling algorithm provides the filtering and the filtering by the DAC itself can be less steep.

 

[pma]

To prevent mirror images in the audible domain.

Well, yes, but - to prevent mirror images above Fs/2, in case of DAC. Like here, 44.1kHz - 21.5kHz makes 22.6kHz mirror. We do not want possible intermodulation 22.6 - 21.5 = 1.1kHz in the amplifier.

In case of ADC, alias would fall below Fs/2.

 

[Mulder]

To prevent mirror images in the audible domain.
The upsampling is needed because there are less bits available to work with (usually somewhere between 4 and 6) .
So the filter MUST be steep. Or should I say should be steep.
When one uses software upsampling the upsampling algorithm provides the filtering and the filtering by the DAC itself can be less steep.

Does this mean there is a potential benefit from software oversampling prior the DAC?

 

[solderdude]

Does this mean there is a potential benefit from software oversampling prior the DAC?

Yes, there certainly can be when the DAC in question has no or a very poor reconstruction filter and 44.1 or 48kHz files are used.
Mirror images (above the audible range) can alias down into the audible range and this is not harmonic-al to the music.
This aliasing can be caused by non linear behavior (distortion) in the speakers or even amplifiers, even more so with class-D.

 

[danadam]

However, my point is that nowadays upsampling is used within the modern DAC's. Therefore, just to avoid imaging the filters don't need to be steep anymore. But you still see steep filters in the designs.. So my question is, why?

There are two filters in the process, one digital and one analog. The digital filter is part of upsampling and it is steep. The analog filter is part of actual conversion to analog and it can be shallow. The measurements show their combined effect. So when steepness is judged, it is the upsampling digital filter that is being judged, not the analog one.

 

[Lambda]

just to avoid imaging the filters don't need to be steep anymore. But you still see steep filters in the designs.. So my question is, why?

The filtter needs to have the same steep response but with an upsampled signal the filter can be digital.

Still I see in all reviews filters with a cut-off at approx. 22 kHz.

because Amir is only testing with 44.1k

 

[Endibol]

There are two filters in the process, one digital and one analog. The digital filter is part of upsampling and it is steep. The analog filter is part of actual conversion to analog and it can be shallow. The measurements show their combined effect. So when steepness is judged, it is the upsampling digital filter that is being judged, not the analog one.

Thanks a lot. I am not sure if I already fully understand it, but you bring me on the “right track”..