The comedy of some Hi res recordings

[Martin_320]

I haven't said it's a "problem". I'm just describing the characteristics and benefits of what happens when higher sample rates are used. ie you get this nice wide transition band. Which is why I like my music preferably to be at high sample rates; and high bit-depths, because I want to have my AVP's 32-bit DSPs fed with as much data to work with.
No audio engineer worth his salt would even dream of mixing or mastering his next chart hit in only 16 bit resolution!
I'm applying the same logic here with regard to the DSPs which are very much active in my Marantz AVP.

 

[MRC01]

... Nevertheless, 44.1kHz is adequate bandwidth for an anti-imaging filter that has flat phase and amplitude characteristics within the entire audio band, and there is no shortcoming with regard to implementing such filters on cheaper DACs (indeed one generally needs to pay a fancy DAC manufacturer quite a bit of money to intentionally f*** this up, IME).

You'll get something like this:
Here is one for a DAC with several filter choices.

Look carefully at the graph above and it disproves your point. 44.1 kHz is not adequate bandwidth for an anti-imaging filter that has flat phase & amplitude within the entire audio band, and runs in real-time on limited hardware (a typical DAC chip). The standard filter shown in that graph doesn't fully attenuate at 22.05 like it should; instead, it is not fully attenuated until 24.1. This is actually common across most of the DACs measured here. Why is that? Why are all these DACs breaking the Nyquist rule? (and only breaking it at 44.1 kHz sampling?) Because at 44.1 kHz sampling, the transition band is only 20k to 22.05 which is too narrow to implement a filter that is flat in phase & amplitude and runs in real-time. And why do they pick 24.1 as the stopband (instead of 22.05)? Because alias frequencies are mirror-imaged around Nyquist, this keeps any supersonic noise that the filter passes, above 20 kHz so it should be inaudible to humans. Put differently, it is no coincidence that 22,050 (Nyquist) is the exact center between 20,000 (top of passband) and 24,100 (stopband).

One could argue that this is harmless because whatever supersonic noise above Nyquist that this filter passes, is inaudible. That is true in and of itself. However, intermodulation distortion in the electronics downstream from the DAC could create passband distortion from these supersonic frequencies.

At higher sampling rates, these chips don't break the Nyquist rule. They don't have to, because at higher sampling rates the transition band is wider, obviating the need to break the rules.

 

[andreasmaaan]

I haven't said it's a "problem". I'm just describing the characteristics and benefits of what happens when higher sample rates are used. ie you get this nice wide transition band. Which is why I like my music preferably to be at high sample rates; and high bit-depths, because I want to have my AVP's 32-bit DSPs fed with as much data to work with.

I agree it won't be a "problem" unless the images affect downstream components, but it's certainly suboptimal, and would be easily fixed by using a filter with low passband ripple, flat passband phase, and high stopband attenuation (i.e. what I have been referring to in previous posts as a "proper" filter).

 

[Blumlein 88]

Look carefully at the graph above and it disproves your point. 44.1 kHz is not adequate bandwidth for an anti-imaging filter that has flat phase & amplitude within the entire audio band, and runs in real-time on limited hardware (a typical DAC chip). The standard filter shown in that graph doesn't fully attenuate at 22.05 like it should; instead, it is not fully attenuated until 24.1. This is actually common across most of the DACs measured here. Why is that? Why are all these DACs breaking the Nyquist rule? (and only breaking it at 44.1 kHz sampling?) Because at 44.1 kHz sampling, the transition band is only 20k to 22.05 which is too narrow to implement a filter that is flat in phase & amplitude and runs in real-time. And why do they pick 24.1 as the stopband (instead of 22.05)? Because alias frequencies are mirror-imaged around Nyquist, this keeps any supersonic noise that the filter passes, above 20 kHz so it should be inaudible to humans. Put differently, it is no coincidence that 22,050 (Nyquist) is the exact center between 20,000 (top of passband) and 24,100 (stopband).

One could argue that this is harmless because whatever supersonic noise above Nyquist that this filter passes, is inaudible. That is true in and of itself. However, intermodulation distortion in the electronics downstream from the DAC could create passband distortion from these supersonic frequencies.

At higher sampling rates, these chips don't break the Nyquist rule. They don't have to, because at higher sampling rates the transition band is wider, obviating the need to break the rules.

The filter shown is common. It is called a half band filter by some. It is a cost saving measure. The assumption is you can do half the filter on the ADC end and half on the DAC end with the same result of a more complex filter. Doing the test where you look at a digitally created file played back and recorded at a higher rate you aren't getting the same steepness as if the file being played back had already gone thru the same filter once when it went thru an ADC. The result still isn't full attenuation by 22.05 khz, but it isn't very far above that mark. Most DACs show the same pattern at all sample rates.

 

[Martin_320]

I agree it won't be a "problem" unless the images affect downstream components, but it's certainly suboptimal, and would be easily fixed by using a filter with low passband ripple, flat passband phase, and high stopband attenuation (i.e. what I have been referring to in previous posts as a "proper" filter).

I see now you bring "ripple" into the discussion ;-)
Unfortunately I doubt that there is any one single filter which excels in all of these properties simultaneously, ie:
"low passband ripple" AND "flat passband phase" AND "high stopband attenuation.
This is exactly why companies like Sabre and AKM offer at least three types of filters in their DACs. So you can take your pick on which of the aforementioned properties you wish to prioritise.

ps. The simple way to get around the filters' respective compromises is use much higher sample rates, all the way from the original studio recording A-to-D, thru to the playback D-to-A process, with no resampling happening at any point along the way.

 

[andreasmaaan]

I see now you bring "ripple" into the discussion ;-)

I didn't just bring it into the discussion now. I simply changed the language I'm using to describe it (previously I had referred to "flat passband response" or similar).

Unfortunately I doubt that there is any one single filter which excels in all of these properties simultaneously, ie:
"low passband ripple" AND "flat passband phase" AND "high stopband attenuation.

In AKM DACs, for example, this filter is known as “sharp roll-off”. It has passband ripple under +/-0.01dB (an order of magnitude below what could be audible), constant group delay, and 100dB of stopband attenuation.

I forget what ESS calls their version.

This is exactly why companies like Sabre and AKM offer at least three types of filters in their DACs. So you can take your pick on which of the aforementioned properties you wish to prioritise.

There are two reasons they do this. One reason is to cater to audiophile nervosa. The second is that linear-phase filters introduce more latency, which is fine for most purposes, but may be undesirable in some applications.

In the case of AKM DACs, IIRC the lower-latency filter is known as “short delay sharp roll-off. It has flat passband amplitude response and high stopband attenuation, but non-constant group delay.

 

[Martin_320]

Further to above, if you want to know what that

In AKM DACs, for example, this filter is known as “sharp roll-off”. It has passband ripple under +/-0.01dB (an order of magnitude below what could be audible), constant group delay, and 100dB of stopband attenuation.

AKM specifically offer the "Slow Roll-off" filter to avoid the passband ripple effect.
Just compare for yourself AKM's respective plots at the bottom of page 14 ("Sharp Roll-off Filter Passband Ripple") and bottom of page 16 ("Slow Roll-off Filter Passband Ripple") respectively...
https://www.akm.com/content/dam/doc.../audio-dac/ak4490eq/ak4490eq-en-datasheet.pdf

 

[andreasmaaan]

Further to above, if you want to know what that


AKM specifically offer the "Slow Roll-off" filter to avoid the passband ripple effect.
Just compare for yourself AKM's respective plots at the bottom of page 14 ("Sharp Roll-off Filter Passband Ripple") and bottom of page 16 ("Slow Roll-off Filter Passband Ripple") respectively...
https://www.akm.com/content/dam/doc.../audio-dac/ak4490eq/ak4490eq-en-datasheet.pdf

Yes, this is their nod to audiophile nervosa. Their sharp filter’s passband ripple is <0.01dB. That’s more than ten times lower than any human can hear. The slow filter has no audible benefit. All it does is introduce imaging, which may well lead to audible issues in some circumstances.

 

[MRC01]

Different manufacturers implement "slow rolloff" in different ways. This is especially evident at high sampling rates (above 44.1).

For example, the Wolfson/Cirrus Logic WM8741 has 5 built-in filters. Some fast, some slow. For this chip, slow rolloff is fully attenuated at Nyquist, but it starts the transition band at a lower frequency, yet still above 20 kHz. In my view, this is a "proper" implementation, taking advantage of the wider transition band for cleaner passband performance, yet still blocking all supersonic noise. The only downside is starting the transition band earlier (at a lower frequency), yet the transition band still starts above 20 kHz so it's attenuating frequencies we can't hear anyway.

Put differently: consider a filter for sampling at 96 kHz. There is no point to passing 40 kHz without attenuation. We can't hear that. Most dogs can't even hear that. You might as well start the transition band just above 20 khz to gradually taper to full attenuation by 48 kHz. That's what the WM8741's slow filters do at high sampling frequencies.

For the AK4490 you linked above, slow rolloff does not fully attenuate by Nyquist. These filters are down only about -6 dB at Nyquist. This allows supersonic noise to pass making it an "improper" filter. Why they would design it like this makes no sense to me because at high sampling frequencies where Nyquist is at least an octave above 20 kHz, you can use a gradual slope yet still fully attenuate by Nyquist.

 

[Martin_320]

Yes, this is their nod to audiophile nervosa. Their sharp filter’s passband ripple is <0.01dB. That’s more than ten times lower than any human can hear. The slow filter has no audible benefit. All it does is introduce imaging, which may well lead to audible issues in some circumstances.

It is not correct to say that a slow filter choice "introduces" imaging.
Rather, the images are already there as a by-product of the prior 'sample-and-hold' process.

 

[MRC01]

Due to differences in filter implementations, it would be more accurate to talk about filters that don't fully attenuate by Nyquist. It is commonly believed that only "slow" filters have this issue. But some slow filters (like those on the WM8741) don't have this issue. And some "fast" filters do have this issue -- like those whose graph is posted above (where stopband at 44.1k sampling is 24.1k).

So it's not really about "fast" vs. "slow". It's about whether filter fully attenuates by Nyquist.

 

[Martin_320]

So it's not really about "fast" vs. "slow". It's about whether filter fully attenuates by Nyquist.

This point is moot anyway. With real music sampled at 96kHz rate, there is practically no supersonic audio content beyond 30kHz. Which means there is a 18kHz-wide gap of nothing-at-all between 30kHz and the Nyquist (48kHz). So with real music there can be no significant reflected "image" immediately above Nyquist (48kHz) -- until you get to around 66kHz in the spectrum -- and by that point even the slowest of slow filters would have pushed it down to virtually nothing.

 

[andreasmaaan]

It is not correct to say that a slow filter choice "introduces" imaging.
Rather, the images are already there as a by-product of the prior 'sample-and-hold' process.

Slow filters that don’t adequately attenuate by Nyquist fail to avoid imaging.

 

[Martin_320]

Slow filters that don’t adequately attenuate by Nyquist fail to avoid imaging.

Depends what you mean by "adequately".
See my previous post ...

 

[andreasmaaan]

Depends what you mean by "adequately".
See my previous post ...

I saw your previous post. I don’t disagree.

But it doesn’t contradict my position that you’ve been arguing against this entire time: namely, that 44.1kHz is a perfectly adequate sampling rate for a filter with constant group delay and flat amplitude response in the passband and adequate stopband attenuation.

I’ve made this point a number of times and provided evidence for it, and can’t see how you could still possibly disagree with it, so I think I’ll leave it there for now

 

[MRC01]

... But it doesn’t contradict my position that you’ve been arguing against this entire time: namely, that 44.1kHz is a perfectly adequate sampling rate for a filter with constant group delay and flat amplitude response in the passband and adequate stopband attenuation. ...

As I mentioned above, and as seen in the graph that @Blumlein 88 posted, 44.1 kHz is not quite adequate. Its filter transition band is so narrow that most DAC chips break the rules and stretch it (stopband above Nyquist) in order to achieve ideal amplitude and phase response in the passband.

 

[mansr]

A little imaging just above 22.05 kHz isn't a problem. Firstly, the frequencies are inaudible. Secondly, the imaged frequencies have little content in the first place. Thirdly, downstream devices are usually well-behaved at least up to 25 kHz, so IMD-induced artefacts in the audible range are minimal. The output may not be strictly by the book, but it really doesn't matter.

 

[Blumlein 88]

Would have been nice to have standardized on JJ's suggestions. 65 khz sampling. Flat response to 30 khz, and a 5 khz transition band.

 

[andreasmaaan]

As I mentioned above, and as seen in the graph that @Blumlein 88 posted, 44.1 kHz is not quite adequate. Its filter transition band is so narrow that most DAC chips break the rules and stretch it (stopband above Nyquist) in order to achieve ideal amplitude and phase response in the passband.

I stand corrected, thanks @MRC01. I was mistakenly of the view that the most common chips in well-performing affordable DACs (e.g. AKM velvet sound, ESS90xx) did not do this. A second look at a number of datasheets confirms I was wrong in the case of AKM (I can't find info on ESS).

Also, now that I look at these spec sheets in more detail, it seems that the same stopband/sampling rate ratio is present at all sampling rates, not just 44.1kHz. For example, AKM velvet series chips all place the stopband of their "sharp" roll-off at approximately 0.545*fs for all sampling rates, and approximately 0.89*fs for their "slow" roll-off filters.

In other words, imaging is unavoidable with these chips, regardless of sampling rate (albeit better with the faster roll-off filters).

 

[MRC01]

... it seems that the same stopband/sampling rate ratio is present at all sampling rates, not just 44.1kHz. For example, AKM velvet series chips all place the stopband of their "sharp" roll-off at approximately 0.545*fs for all sampling rates, and approximately 0.89*fs for their "slow" roll-off filters.
In other words, imaging is unavoidable with these chips, regardless of sampling rate (albeit better with the faster roll-off filters).

This seems to vary from chip to chip. With the aforementioned WM8741, the stopband is 0.5fs or less at high sampling rates, for both slow and fast rolloff. With this chip, stopband above 0.5fs is unique to low sampling rates (44/48).

That said, a stopband of 24.1k at 44.1k sampling (e.g. stretched beyond Nyquist) is common to many/most of the DACs that Amir reviews here.