Is Hi-Res Audio better because filter is outside audible range?

[33AndAThird]

Hi all,

I listened to the Darko Audio podcast 10 hi-fi myths busted with Peter Comeau, and have questions about the first myth "Hi-res audio sounds better because of its ultrasonic content". I figured this would be the best place to get some straight, evidence based answers I know there will be some eye-rolls as to the source, but my questions stem from wanting to understand.

Peter Comeau starts out by confirming what we all know to be true, which is that anything above 20KHz is inaudible and the proposition that one can get a better musical experience by reproducing frequencies above 20KHz is nonsense. He then went on to say that hi-res does indeed sound better not because of reproducing higher frequencies but that the filter is well into the ultrasonic region and hence not audible. He gives the example of early CD players having a brick-wall filter at 20KHz and that being detrimental to performance.
My questions are:
1. Why is a filter required in the first place? I superficially understand Nyquist, and that a 10KHz signal will also reproduce a 34.1KHz alias signal when converting back to analog, but why do we care? If we can't hear it, and likely the equipment can't reproduce it anyway, why do we need to bother with a reconstruction filter that shops off everything above 20KHz?
2. On the assumption that the filter is required, why is a brickwall filter range a bad thing compared to others? Peter talks about the issue being in the time domain, and that the 20KHz brickwall filter of the early CD days had both pre and post echoes. He said that this can be viewed on an oscilloscope and the evidence is online (I couldn't find it but its entirely possible I'm using the wrong terminology). Is there evidence to suggest a brickwall is a bad thing, and why do we get artifacts in the time domain?
3. The summary given is "By having a higher sample rate, your filter is moved further away from the upper limit of the audible band, so it doesn't trouble as much the time-domain response of the signal". Presumably this means that any filter, brickwall or not, has a negative impact to some degree and we should avoid them in the audible range. Is there evidence for this?

Any help in answering my questions is much appreciated! Thanks

 

[Audiofire]

Peter talks about the issue being in the time domain, and that the 20KHz brickwall filter of the early CD days had both pre and post echoes. He said that this can be viewed on an oscilloscope and the evidence is online (I couldn't find it but its entirely possible I'm using the wrong terminology).

Images from this thread: https://www.audiosciencereview.com/...sampling-dac-to-buy.43870/page-2#post-1558974

The summary given is "By having a higher sample rate, your filter is moved further away from the upper limit of the audible band, so it doesn't trouble as much the time-domain response of the signal". Presumably this means that any filter, brickwall or not, has a negative impact to some degree and we should avoid them in the audible range. Is there evidence for this?

Yes, but it depends on the Nyquist frequency in relation to a given signal as well as engineering implementations.

 

[lemmy_collins]

1st things 1st: aliasing can occur when sampling (ie recording) if the highest frequency in the analog content is more than half the sampling freq, thus by sampling at 44.1kHz, an unfiltered 32.05kHz analog freq will alias into a 10kHz digital one... I can't see how an aliasing process could occur on the digital to analog side (but i could be wrong).

Another details: a "brickwall filter" can't be implemented in reality, any real filter "gradually" rejects out-of-interest frequencies, the slop depending on the order of the filter: 1st order filters reject by 6dB per octave, thus given a low-pass filter cutting at 20kHz (generally given at -3dB), a 1st order filter attenuation will be of -9dB at 40kHz, a 2nd order -15dB and so on... But it is true that high order filter are hard to do because they quickly become unstable (true for analog and numeric filters).

Now the dephasing (another way to call time delaying _depending_ on the frequencies) occurs only when the filter starts to reject and the highest the order is, the more it dephases (delays)... But the higher the freq are in the rejected band, the less we care since the freq are more and more attenuated.

All that said, making the Nyquist frequency higher to ease the anti-aliasing filter was a thing until analog to digital converters were sigma-delta based (since more than 20y...)

 

[lemmy_collins]

Images from this thread: https://www.audiosciencereview.com/...sampling-dac-to-buy.43870/page-2#post-1558974

Yes, but it depends on the Nyquist frequency in relation to a given signal as well as engineering implementations.

OK we're talking about reconstruction filters... not sure how to translate what I know from ADC to DAC world but there's an analogy.

The "sinus-cardinal" in that context could be a "brick wall" filter since it is time bounded once the value are less that 1 bit... But the more there's coefficients to the filter, the more CPU hungry it is... Maybe there's tricks to approach it, i've seen pure sinus at sampling rate / 2 with RME DAC there's more than 15y ago...

 

[rationaltime]

My questions are:
1. Why is a filter required in the first place? I superficially understand Nyquist, and that a 10KHz signal will also reproduce a 34.1KHz alias signal when converting back to analog, but why do we care? If we can't hear it, and likely the equipment can't reproduce it anyway, why do we need to bother with a reconstruction filter that shops off everything above 20KHz?

Any help in answering my questions is much appreciated! Thanks

Let's talk about sampling.
As lemmy_collins has mentioned at the start, the issue is the aliasing of
frequency components above the Nyquist limit down into the audible
frequency range. To emphasize that, down conversion would become
an issue if the signal is not band limited.

 

[Blumlein 88]

Let's talk about sampling.
As lemmy_collins has mentioned at the start, the issue is the aliasing of
frequency components above the Nyquist limit down into the audible
frequency range. To emphasize that, down conversion would become
an issue if the signal is not band limited.

Only on the ADC end. There is no aliasing down upon playback. If it was recorded in the file, you cannot get rid of it. On playback on the DAC end you only have imaging above the nyquist rate. It is possible that if high enough in level that could modulate back down via IMD into the audible range, but that nearly never is the case.

 

[Blumlein 88]

Here is a test by JA at Stereophile of a Mytek Brooklyn II. JA feeds a high level 19.1 khz tone and looks wideband at the output. Showing the filter response using white noise. If there is imaging it will show up as a tone at 25 khz. Here you see it is around - 50 db at 25 khz.

Mytek Digital Brooklyn Bridge II Roon Core preamplifier Measurements

Sidebar 3: Measurements

www.stereophile.com

Here is the same measurement of a Weiss Helios that has a steeper filter and it suppresses the 25 khz imaging tone as you would expect.

Weiss Engineering Helios D/A processor Measurements

Sidebar 3: Measurements

www.stereophile.com

It is remotely possible that the "ringing" might be heard by some small number of people in unusual circumstances. You will notice those charts of "ringing" show ringing at the 22,050 hz rate. You aren't likely to hear that. Few if any people can. It is true 88.2 or 96 khz rates would eliminate the possibility. J_J if you know who he is (was at Bell Labs) wanted a 64 khz sample rate in the early days. This would have made the filter a non-issue by allowing a wider stop band between the sampling Nyquist rate and our hearing limit at 20 khz. As most adults by 30 don't have 20 khz hearing it cannot matter to many people.

 

[manisandher]

It's easy to show that brickwall filters don't do anything untoward.

Take a hires file, say 24/192, with lots of content above 24kHz. 'Downsample' to 48kHz using a brickwall anti-alias filter, and then 'upsample' back to 192kHz using a brickwall anti-imaging filter. Now compare the 'down-/up-sampled' file with the original. It will null to <-210dB in the passband (i.e. below 24kHz).

Good enough for me.

Mani.

 

[manisandher]

But this is what all anti-imaging filters should really look like:

(Stereophile measurements of Ferrum Wandla D/A.)

Full attenuation before Nyquist.

Mani.

 

[KSTR]

@33AndAThird , you might want to have some fun listening to this:

How DAC reconstruction filters would sound like if we were bats (listening examples)

There is a lot of discussion on how different DAC reconstruction filters sound like, besides the pure technical aspects which are easily explained and visualized with frequency response plots where the imaging -- mirrored frequencies above half the sample rate -- can be seen (if any), as well as...

www.audiosciencereview.com

Then we should ask ourselves how much of the effects might still be relevant at original sample rates. That will mostly depend on the conditions of one's ears (age and abuse) but also on listening skills.

 

[Kwesi]

I have attached some paper from german company Klein + Hummel, now Neumann (Yes, the Neumann company with the highly regarded studio monitors...)

Translation via deepl:

"Hearing above 20 kHz

For several decades, science has assumed that the upper limit of the human hearing frequency range is at best around 20 kHz; for many older adults, the "upper limit frequency" is only 16 kHz or even lower.
Nevertheless, recent listening tests with music material conducted as part of the 96/192 kHz discussion have revealed differences between signals limited by the transmission chain up to 20 kHz and those reproduced via a transmission chain with a wider frequency range.
Against this background, the two scientists Ashihara Kaoru and Kiryu Shogo from the Japanese ETL Institute carried out various experiments and listening tests, the results of which were presented to experts at the 7th AES Convention (Paper No. 5401). According to this, the existence of actually inaudible frequencies is perceived, among other things, when they are reproduced together with other complex, non-sinusoidal signals via a non-linear transmission system. Due to these non-linear
transmission characteristics, the higher frequency signals are demodulated into the audible frequency range. Transmission elements of the audio chain in which the non-linear
transmission behavior is relatively pronounced are, for example, the loudspeakers through which the music signal is reproduced.

The above facts were demonstrated by the following experiment:
At a distance of approx. 220 cm from the (normally hearing) test subject, two loudspeakers are mounted directly above each other in such a way that they have identical transfer functions at the listening position. In the first part of the test, only one of the two loudspeakers is used: A harmonic frequency spectrum is reproduced via this speaker, which is limited upwards to 35 kHz and whose fundamental wave has been filtered out. This spectrum was then supplemented by a tone with a frequency of 31.5 kHz, which was
pulsed with an additional frequency of 2 Hz to make it stand out better from the rest of the signal. All test subjects heard the difference between the two signals, although the frequency of the pulsed tone was far above their hearing frequency range. In a subsequent FFT analysis of the acoustic signal at the listening position, an additional component at 3.5 kHz was detected in the second test signal (with pulsed 31.5 kHz sound):
This is a (clearly audible) intermodulation distortion caused by the non-linear behavior of the loudspeakers. caused by non-linear behavior of the speakers.

In the second part of the experiment, both speakers are now used: One reproduces the test spectrum (without the 31.5 kHz tone), while the second is used exclusively to transmit the 31.5 kHz tone (again pulsed at 2 Hz). This eliminates the non-linear loudspeaker characteristics in relation to the test tone. In this case, none of the test subjects could make out a difference in the performance with the 31.5 kHz tone switched on or off, even when its sound pressure level was above 80 dB. The spectra measured later at the listening position differed only in the existence of the 31.5 kHz tone; no component was found at 3.5 kHz.

Conclusion: Differences that can certainly be heard in listening tests over transmission chains with different upper cut-off frequencies (> 20 kHz) can be attributed to additional low-frequency intermodulation products, including those of the loudspeakers, which occur when these are excited with higher-frequency (useful) signals. Meanwhile, the high-frequency signals belonging to the original performance are not perceived by humans!"


Personal experience with the "tweeter modulation effect":

I have tested the effect of passive notch filters in series of tweeters, namely Seas DXT and Seas TAC/GB, to surpress the aluminum dome breakup resonance ~27kHz.
(Note: My heraring range ends above ~16kHz...)

Then giving it sine sweeps with 96kHz sampling frequency via Arta measurement SW, so full excitement up to 48kHz fully in the resonance range. Listening repeated to the sweeps, switching a bypass to the passive notch on and off. With bypassed notch, I percieved a sharp note in the reverberant tone that dissapears when the notch is active.

Blind test? When you might test it yourself you will laugh that you ever asked, so obvious the effect is...

Of course, during normal music playback operation in my system the DAC cuts off everything >20kHz, so the dome breakup gets never excited with full level as in my test, but only from lower frequencies distortion that is magnitudes lower.


Perhaps the difference of the mentioned filters is only their frequency response without further phase or impulse behaviour effect, that leads to different excitement of the tweeter >20kHz with high sampling rate music?

 

[33AndAThird]

Thanks very much everyone for the detailed responses, lots to get into here.

In addition to replies here, I found this page very useful in helping me understand why filters in the frequency domain cause issues (rininging) in the time domai, if anyone is interested.

@33AndAThird , you might want to have some fun listening to this:

@KSTR this was super interesting thank you. The NOS filter was indeed very obvious, though I did have trouble distinghuishing the minimum phase and linear phase filter. Very intersting to hear what the mirror effects sound like at lower frequencies though.

t's easy to show that brickwall filters don't do anything untoward.

Take a hires file, say 24/192, with lots of content above 24kHz. 'Downsample' to 48kHz using a brickwall anti-alias filter, and then 'upsample' back to 192kHz using a brickwall anti-imaging filter. Now compare the 'down-/up-sampled' file with the original. It will null to <-210dB in the passband (i.e. below 24kHz).

Good enough for me.

Also very intersting - so does this imply that the exporation of various different filters in DACS is for no audible benefit?

Let's talk about sampling.
As lemmy_collins has mentioned at the start, the issue is the aliasing of
frequency components above the Nyquist limit down into the audible
frequency range. To emphasize that, down conversion would become
an issue if the signal is not band limited.

So my understanding in the example (and realise now I should have provided this context), is that when creating early CD's studio engineers used analog filters to cut out any signal above 20KHz. So we can assume that there was no 20KHz+ signal sampled.
The problem referred to is the reconstruction filter used in early CD players which implemented a 20KHz+ brickwall. I understand that sub 20KHz frequeencies created 20KHz+ aliases on reconstruction but I still don't understand why we should care about this? Why do we bother with a brickwall filter (or any other filter) if it introduces other artifacts like riniging?

 

[Blumlein 88]

I have attached some paper from german company Klein + Hummel, now Neumann (Yes, the Neumann company with the highly regarded studio monitors...)

Translation via deepl:

"Hearing above 20 kHz

For several decades, science has assumed that the upper limit of the human hearing frequency range is at best around 20 kHz; for many older adults, the "upper limit frequency" is only 16 kHz or even lower.
Nevertheless, recent listening tests with music material conducted as part of the 96/192 kHz discussion have revealed differences between signals limited by the transmission chain up to 20 kHz and those reproduced via a transmission chain with a wider frequency range.
Against this background, the two scientists Ashihara Kaoru and Kiryu Shogo from the Japanese ETL Institute carried out various experiments and listening tests, the results of which were presented to experts at the 7th AES Convention (Paper No. 5401). According to this, the existence of actually inaudible frequencies is perceived, among other things, when they are reproduced together with other complex, non-sinusoidal signals via a non-linear transmission system. Due to these non-linear
transmission characteristics, the higher frequency signals are demodulated into the audible frequency range. Transmission elements of the audio chain in which the non-linear
transmission behavior is relatively pronounced are, for example, the loudspeakers through which the music signal is reproduced.

The above facts were demonstrated by the following experiment:
At a distance of approx. 220 cm from the (normally hearing) test subject, two loudspeakers are mounted directly above each other in such a way that they have identical transfer functions at the listening position. In the first part of the test, only one of the two loudspeakers is used: A harmonic frequency spectrum is reproduced via this speaker, which is limited upwards to 35 kHz and whose fundamental wave has been filtered out. This spectrum was then supplemented by a tone with a frequency of 31.5 kHz, which was
pulsed with an additional frequency of 2 Hz to make it stand out better from the rest of the signal. All test subjects heard the difference between the two signals, although the frequency of the pulsed tone was far above their hearing frequency range. In a subsequent FFT analysis of the acoustic signal at the listening position, an additional component at 3.5 kHz was detected in the second test signal (with pulsed 31.5 kHz sound):
This is a (clearly audible) intermodulation distortion caused by the non-linear behavior of the loudspeakers. caused by non-linear behavior of the speakers.

In the second part of the experiment, both speakers are now used: One reproduces the test spectrum (without the 31.5 kHz tone), while the second is used exclusively to transmit the 31.5 kHz tone (again pulsed at 2 Hz). This eliminates the non-linear loudspeaker characteristics in relation to the test tone. In this case, none of the test subjects could make out a difference in the performance with the 31.5 kHz tone switched on or off, even when its sound pressure level was above 80 dB. The spectra measured later at the listening position differed only in the existence of the 31.5 kHz tone; no component was found at 3.5 kHz.

Conclusion: Differences that can certainly be heard in listening tests over transmission chains with different upper cut-off frequencies (> 20 kHz) can be attributed to additional low-frequency intermodulation products, including those of the loudspeakers, which occur when these are excited with higher-frequency (useful) signals. Meanwhile, the high-frequency signals belonging to the original performance are not perceived by humans!"


Personal experience with the "tweeter modulation effect":

I have tested the effect of passive notch filters in series of tweeters, namely Seas DXT and Seas TAC/GB, to surpress the aluminum dome breakup resonance ~27kHz.
(Note: My heraring range ends above ~16kHz...)

Then giving it sine sweeps with 96kHz sampling frequency via Arta measurement SW, so full excitement up to 48kHz fully in the resonance range. Listening repeated to the sweeps, switching a bypass to the passive notch on and off. With bypassed notch, I percieved a sharp note in the reverberant tone that dissapears when the notch is active.

Blind test? When you might test it yourself you will laugh that you ever asked, so obvious the effect is...

Of course, during normal music playback operation in my system the DAC cuts off everything >20kHz, so the dome breakup gets never excited with full level as in my test, but only from lower frequencies distortion that is magnitudes lower.


Perhaps the difference of the mentioned filters is only their frequency response without further phase or impulse behaviour effect, that leads to different excitement of the tweeter >20kHz with high sampling rate music?

While all of this is true think about it. In real life you don't have those non-linear effects. So while higher sample rates can be detected as different for playback due to speaker non-linearity that was not there in the original. So the higher sample rate is creating a possibly audible effect, not reproducing the music at higher fidelity. It is producing it at lower fidelity which sounds different. A case for limiting response rather than extending it.

 

[MRC01]

There is another benefit to "high res" which is you get a wider transition band, which makes it easier to implement a proper filter. By proper filter I mean from an engineering perspective. Whether that is audible is a different question.

CD at 44.1 kHz is not sufficient to implement a proper filter in real-time with the computation power available in typical DAC chips, as you can see in just about every DAC review on this site. The transition band of 20 k to 22.05 k is too narrow, so they either start attenuating in the passband or they stretch the transition band past Nyquist and don't achieve full attenuation by 22.05 kHz - typically at 24.1 kHz. This allows some HF noise to leak through, but due to way aliasing works, the common stopband of 24,100 is no coincidence. It means any HF aliasing will be above 20 kHz. That is, 22,050 - (24,100 - 22,050) = 20,000 Hz.

However, at 48 kHz and higher most good DACs do have a "proper" filter - full attenuation by Nyquist without any passband attenuation. So it only takes a slightly higher sampling frequency to get a wide enough transition band to implement a proper filter.

 

[danadam]

The transition band of 20 k to 22.05 k is too narrow, so they either start attenuating in the passband or they stretch the transition band past Nyquist and don't achieve full attenuation by 22.05 kHz - typically at 24.1 kHz. This allows some HF noise to leak through, but due to way aliasing works, the common stopband of 24,100 is no coincidence. It means any HF aliasing will be above 20 kHz. That is, 22,050 - (24,100 - 22,050) = 20,000 Hz.

AFAIU...
Aliasing happens only in AD conversion, so the above description applies only to ADCs. DACs on the other hand produce images, so the consequence of stretching the transition band to 24 kHz should be just a bit more HF noise in the output. The HF noise can potentially cause IMD down the line but that's different from aliasing.

 

[Audiofire]

Aliasing happens only in AD conversion, so the above description applies only to ADCs.

There are ways to not use a low-pass filter when downsampling an already digitized audio file, which can cause aliasing.

 

[dualazmak]

Just for your possible reference and interest, @amirm and we have intensively "measured" and discussed about possible UHF (ultra-high frequency) undesirable "noises" often included in poorly QC-ed high-resolution music tracks especially in DSD format.

Consequently, I, at least myself, decided to have -48 dB/Oct low-pass (high-cut) DSP filters at 25 kHz for midrange, tweeter and super-tweeter Fq zones and drivers.

Even though my DSP EKIO can process upto in 192 kHz 24 bit, this issue resulted in converting on-the-fly all the digital music tracks into 24 bit 88.2 kHz or 96.0 kHz PCM by JRiver MC which is high-bitrate enough since I have high-cut filters at 25 kHz in system-wide DSP "EKIO".

Spoiler

On my project thread...
I (we) fully discussed and evaluated the UHF (ultra-high frequency) noise issue in poorly QC-ed HiRes music tracks including DSD formats;
- "Near ultrasound - ultrasound" ultra-high frequency (UHF) noises in improperly engineered/processed HiRes music tracks, and EKIO's XO-EQ configuration to cut-off such noises: #362-#386, #518

- Summary of rationales for "on-the-fly (real-time)" conversion of all music tracks (including 1 bit DSD tracks) into 88.2 kHz or 96 kHz PCM format for DSP (XO/EQ) processing: #532

- The latest system setup of my DSP-based multichannel multi-SP-driver multi-amplifier fully active audio rig as of August 3, 2023: #774

 

[manisandher]

CD at 44.1 kHz is not sufficient to implement a proper filter in real-time with the computation power available in typical DAC chips, as you can see in just about every DAC review on this site.

See my post #9: https://www.audiosciencereview.com/...r-is-outside-audible-range.53107/post-1919174

Mani.

 

[AJM1981]

I am glad Peter Comeau makes clear choices and didn't integrate a "super tweeter" in the "Super" Denton. Basically that one is just a 3-way speaker in large bookshelf format.

 

[KSTR]

There are ways to not use a low-pass filter when downsampling an already digitized audio file, which can cause aliasing.

It's a rare scenario but yes. Any sampling process is prone to aliasing, not matter if the input signal is analog or digital.