The comedy of some Hi res recordings

[Martin_320]

I've skimmed through this entire topic thread and it seems to me that – no offence – most folks here are barking up the wrong tree, ie. looking for justifications in terms of: 'frequency response' (in relation to sample rates) and 'dynamic range' (in relation to bit-depth).

But these are not the main reasons for going beyond 16bits and 44.1kHz.

Amir, in post #22 hinted the real reason that more than 44.1kHz does confer practical benefits, not just in the production studio, but in the consumer/playback part of the audio chain:

I quote him: "Because it allows more headroom ... High sample rate for example avoids aliasing in non-linear effects."

Specifically, a hirez song (say a FLAC file) at 96kHz will clock your DAC at 96kHz. And when your DAC is being clocked at 96kHz, then your DAC reconstruction filter will start its brickwall filtering much higher up the frequency band. And in doing so any artefacts produced by that brickwall reconstruction filter will occur so far away from the actual human audio band of <20kHz, as to be practically negligible.

The other thing to bear in mind is that your two channel FLAC is in many cases NOT for the final output anymore; rather, the FLAC is just the source data for additional processing stages. Many of us have AVP processors or AVR receivers, so you still benefit from a 24bit word length. "Why?" you may ask.

Well, say if we're upmixing to Dolby Surround (as many of us do), and also applying digital bass-management & room-EQ etc, then the DSPs in your AVP or AVR are doing just as much extra math processing as did the DAW in the studio which mixed the song and exported the FLAC in the first place.

In short, the DSPs in your AVR/AVP are working at least in 32bits, so getting as close to that as possible (ie. by feeding them with 24bit data to work on, as opposed to 16bit) will enable them to work much better -- resulting in much reduced quantization rounding errors at the output from your AVR/AVP's DSP chain. And as for high sample rates, well, these allow your DAC to push any reconstruction filter aliasing artefacts completely out of the way.

So more data (higher sample rates & bit depths at playback) is definitely better.

 

[Sukie]

I've skimmed through this entire topic thread and it seems to me that – no offence – most folks here are barking up the wrong tree, ie. looking for justifications in terms of: 'frequency response' (in relation to sample rates) and 'dynamic range' (in relation to bit-depth).

But these are not the main reasons for going beyond 16bits and 44.1kHz.

Amir, in post #22 hinted the real reason that more than 44.1kHz does confer practical benefits, not just in the production studio, but in the consumer/playback part of the audio chain:

I quote him: "Because it allows more headroom ... High sample rate for example avoids aliasing in non-linear effects."

Specifically, a hirez song (say a FLAC file) at 96kHz will clock your DAC at 96kHz. And when your DAC is being clocked at 96kHz, then your DAC reconstruction filter will start its brickwall filtering much higher up the frequency band. And in doing so any artefacts produced by that brickwall reconstruction filter will occur so far away from the actual human audio band of <20kHz, as to be practically negligible.

The other thing to bear in mind is that your two channel FLAC is in many cases NOT for the final output anymore; rather, the FLAC is just the source data for additional processing stages. Many of us have AVP processors or AVR receivers, so you still benefit from a 24bit word length. "Why?" you may ask.

Well, say if we're upmixing to Dolby Surround (as many of us do), and also applying digital bass-management & room-EQ etc, then the DSPs in your AVP or AVR are doing just as much extra math processing as did the DAW in the studio which mixed the song and exported the FLAC in the first place.

In short, the DSPs in your AVR/AVP are working at least in 32bits, so getting as close to that as possible (ie. by feeding them with 24bit data to work on, as opposed to 16bit) will enable them to work much better -- resulting in much lower rounding errors at the output from your AVR/AVP's DSP chain. And as for high sample rates, well, these allow your DAC to push any reconstruction filter aliasing artefacts completely out of the way.

So more data (higher sample rates & bit depths at playback) is definitely better.

But are we talking about an audible improvement?

 

[Blumlein 88]

I've skimmed through this entire topic thread and it seems to me that – no offence – most folks here are barking up the wrong tree, ie. looking for justifications in terms of: 'frequency response' (in relation to sample rates) and 'dynamic range' (in relation to bit-depth).

But these are not the main reasons for going beyond 16bits and 44.1kHz.

Amir, in post #22 hinted the real reason that more than 44.1kHz does confer practical benefits, not just in the production studio, but in the consumer/playback part of the audio chain:

I quote him: "Because it allows more headroom ... High sample rate for example avoids aliasing in non-linear effects."

Specifically, a hirez song (say a FLAC file) at 96kHz will clock your DAC at 96kHz. And when your DAC is being clocked at 96kHz, then your DAC reconstruction filter will start its brickwall filtering much higher up the frequency band. And in doing so any artefacts produced by that brickwall reconstruction filter will occur so far away from the actual human audio band of <20kHz, as to be practically negligible.

The other thing to bear in mind is that your two channel FLAC is in many cases NOT for the final output anymore; rather, the FLAC is just the source data for additional processing stages. Many of us have AVP processors or AVR receivers, so you still benefit from a 24bit word length. "Why?" you may ask.

Well, say if we're upmixing to Dolby Surround (as many of us do), and also applying digital bass-management & room-EQ etc, then the DSPs in your AVP or AVR are doing just as much extra math processing as did the DAW in the studio which mixed the song and exported the FLAC in the first place.

In short, the DSPs in your AVR/AVP are working at least in 32bits, so getting as close to that as possible (ie. by feeding them with 24bit data to work on, as opposed to 16bit) will enable them to work much better -- resulting in much reduced quantization rounding errors at the output from your AVR/AVP's DSP chain. And as for high sample rates, well, these allow your DAC to push any reconstruction filter aliasing artefacts completely out of the way.

So more data (higher sample rates & bit depths at playback) is definitely better.

I know it is being picky, but reconstruction filters imaging. They are anti-imaging filters. Aliasing occurs during analog to digital conversion, not digital to analog. So anti-aliasing filters are part of converting analog to digital.

Imaging wouldn't show up in the audible band. It can however via IMD cause signals to show up in the audible band.

 

[GDK]

Specifically, a hirez song (say a FLAC file) at 96kHz will clock your DAC at 96kHz. And when your DAC is being clocked at 96kHz, then your DAC reconstruction filter will start its brickwall filtering much higher up the frequency band. And in doing so any artefacts produced by that brickwall reconstruction filter will occur so far away from the actual human audio band of <20kHz, as to be practically negligible.

Yes, but isn‘t that why the DAC chip already oversamples the signal is receives? IIRC the AK4493 chip oversamples by 256 times. Therefore you don’t actually need to feed it a higher frequency signal at all.

 

[andreasmaaan]

Specifically, a hirez song (say a FLAC file) at 96kHz will clock your DAC at 96kHz. And when your DAC is being clocked at 96kHz, then your DAC reconstruction filter will start its brickwall filtering much higher up the frequency band. And in doing so any artefacts produced by that brickwall reconstruction filter will occur so far away from the actual human audio band of <20kHz, as to be practically negligible.

I agree that this may produce audible benefits, but only in cases where a poor reconstruction filter is used. There will be no benefit where a normal filter is used (unless the listener has hearing >20kHz).

In short, the DSPs in your AVR/AVP are working at least in 32bits, so getting as close to that as possible (ie. by feeding them with 24bit data to work on, as opposed to 16bit) will enable them to work much better -- resulting in much lower rounding errors at the output from your AVR/AVP's DSP chain.

Also agree in principle, but again this won't result in audible improvement unless poor-quality dither (or truncation) is used in the first place. Noise-shaped triangular dither (which is industry standard) pushes the noise floor of 16-bit audio below audible levels.

But I agree with you that 16/44 leaves only a very slim buffer. Given the low cost of storage these days, I see no harm in using higher-resolution formats as a way to insure against bad choices in noise-shaping/anti-aliasing/anti-imaging, etc.

 

[Martin_320]

I know it is being picky, but reconstruction filters imaging. They are anti-imaging filters. Aliasing occurs during analog to digital conversion, not digital to analog. So anti-aliasing filters are part of converting analog to digital.
Imaging wouldn't show up in the audible band. It can however via IMD cause signals to show up in the audible band.

Yep -- and same principle applies: With a high DAC sample rate (and thus a higher Nyquist operating frequency), those reconstruction-induced images (which some folks refer to often as 'aliasing') get pushed much higher away from the <20kHz human audible band. This creates a nice clean and very wide "transition band" between the human 20kHz limit and the Nyquist freq - somewhere up beyond 40kHz+ in the spectrum. This means that the DAC's reconstruction filter can do its work well away from your own hearing range.
(I've actually verified this myself by spectrally analysing the analoge output of my AVP at different DAC clock rates.)

In contrast, at 44.1kHz sample rate there is only a very tiny transition band between the human hearing limit (20kHz) and the Nyquist freq (22.05kHz).

 

[andreasmaaan]

With a high DAC sample rate (and thus a higher Nyquist operating frequency), those reconstruction-induced images (which some folks refer to often as 'aliasing') get pushed much higher away from the <20kHz human audible band.

If people refer to it as "aliasing", they are in fact mistaken. There is a key difference not only in the implementation (ADC vs DAC), but also in the effects: A poor anti-aliasing filter will result in aliasing within the audio band, whereas a poor reconstruction filter will result in images above the audio band (which may or may not indirectly lead to in-band distortion depending on downstream equipment).

In contrast, at 44.1kHz sample rate there is only a very tiny transition band between the human hearing limit (20kHz) and the Nyquist freq (22.05kHz).

Tiny, but big enough to allow room for a filter with flat amplitude and phase response within the entire audio band.

 

[Martin_320]

Yes, but isn‘t that why the DAC chip already oversamples the signal is receives? IIRC the AK4493 chip oversamples by 256 times. Therefore you don’t actually need to feed it a higher frequency signal at all.

The "256 times" that you are referring to is likely related to the PWM (Pulse-Width-Modulation) section of the DAC.
(The AK4493, like most general purpose DACs today, is a hybrid DAC which is designed to handle both PCM and DSD streams.)

 

[Martin_320]

...a poor reconstruction filter will result in images ... which may or may not indirectly lead to in-band distortion depending on downstream equipment.

Indeed.

The other important thing is that when the mirrored images are much higher up (with hirez sample rates), starting from 40kHz+ in the spectrum (and I've seen these myself), then the filter doesn't have to be "brick-wall" vertical anymore.
One can use a simpler filter instead (or no filter at all).

 

[andreasmaaan]

Indeed.

The other important thing is that when the mirrored images are much higher up (with hirez sample rates), starting from 40kHz+ in the spectrum (and I've seen these myself), then the filter doesn't have to be "brick-wall" vertical anymore.
One can use a simpler filter instead (or no filter at all).

I think the part of my post that you should have bolded is the following:

a poor reconstruction filter will result in images above the audio band (which may or may not indirectly lead to in-band distortion depending on downstream equipment).

No such problems when the filter is properly implemented.

 

[andreasmaaan]

The "256 times" that you are referring to is likely related to the PWM (Pulse-Width-Modulation) section of the DAC.
(The AK4493, like most general purpose DACs today, is a hybrid DAC which is designed to handle both PCM and DSD streams.)

No, oversampling allows for equivalent bandwidth for the anti-imaging filter to what a higher sample rate would have allowed for in the first place.

For example, a two-times over-sampled 44.1kHz signal and a non-over-sampled 88.2kHz signal are equivalent in terms of allowable bandwidth for the anti-imaging filter.

This is explained quite well here.

 

[Martin_320]

I think the part of my post that you should have bolded is the following:

No such problems when the filter is properly implemented.

Maybe in a super-duper DAC.
But many less expansive DAC filters perform more transparently in a hirez environment -- ie when the available transition band is at least as wide as the audible band.

 

[andreasmaaan]

Maybe in a super-duper DAC.
But many less expansive DAC filters perform more transparently in a hirez environment -- ie when the available transition band is at least as wide as the audible band.

Do you have an example of such a DAC?

IME the more expensive DACs, on average, do this worse (presumably because audiophiles want silly things).

There are also very few DACs these days, cheap or expensive, that don't do oversampling - see my previous post #251.

 

[Martin_320]

No, oversampling allows for equivalent bandwidth for the anti-imaging filter to what a higher sample rate would have allowed for in the first place.

For example, a two-times over-sampled 44.1kHz signal and a non-over-sampled 88.2kHz signal are equivalent in terms of allowable bandwidth for the anti-imaging filter.

This is explained quite well here.

You've missed the point. Internally that DAC in question has a PWM architecture. Yes, it accepts multibit LPCM data, but internally it converts everything to PWM "bitstream". The latter is what the "x256" is referring to.

 

[andreasmaaan]

You've missed the point. Internally that DAC in question has a PWM architecture. Yes, it accepts multibit LPCM data, but internally it converts everything to PWM "bitstream". The latter is what the "x256" is referring to.

Fair enough, I had missed the context of your comment concerning that specific DAC chip.

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).

 

[Martin_320]

Do you have an example of such a DAC?

IME the more expensive DACs, on average, do this worse (presumably because audiophiles want silly things).

There are also very few DACs these days, cheap or expensive, that don't do oversampling - see my previous post #251.

Ok -- I have done tests from the analog-outs of my AVP, with AK4458 DACs, and here's a quick summary of the findings:

Reference file is a 2-ch 96/24 FLAC with real ultrasonic content (ie. not 'noise').

1. Played 'as-is' at full resolution -- Spectral WAV data (from the AVP's analog-outs) shows that it contains real content going out beyond 30kHz -- ultrasonic region. Thereafter diminishing to nothing (no signal, no noise) up to the Nyquist at 48kHz. Images start to appear above 48kHz (but attenuated by slow roll-off filter).

2. Downsample the file in my Studio One DAW to 48kHz. Plays back at 48kHz resolution in AVP -- Spectral playback data shows (from AVP analog outs) that it contains real content going up to 24kHz then a sharp brickwall cut. Then immediately above 24kHz there is clearly an reverse reflection which gradually diminishes due to the slow roll-off filter attenuation.

3. Re-sample the 48kHz file (used in 2. above) back up to 96kHz. Spectral playback analysis from AVP analog outs shows same sharp cut at 24kHz and flat transition band up to 48kHz containing no info. Not much imaging thereafter.

Conclusion: The AKM DACs produce reflection 'images' directly above the Nyquist frequency of the flac being played through it.

Note: Whatever type of "oversampling" may be happening inside an AKM DAC, it is clearly not extending the transition band of LPCM signal beyond that of the original PCM file sample rate. The evidence is on the images which show up clearly just after the original Nyquist freq of the Flac file itself.

 

[andreasmaaan]

Ok -- I have done tests from the analog-outs of my AVP, with AK4458 DACs, and here's a quick summary of the findings:

Reference file is a 2-ch 96/24 FLAC with real ultrasonic content (ie. not 'noise').

1. Played 'as-is' at full resolution -- Spectral WAV data (from the AVP's analog-outs) shows that it contains real content going out beyond 30kHz -- ultrasonic region. Thereafter diminishing to nothing (no signal, no noise) up to the Nyquist at 48kHz. Images start to appear above 48kHz (but attenuated by slow roll-off filter).

2. Downsample the file in my Studio One DAW to 48kHz. Plays back at 48kHz resolution in AVP -- Spectral playback data shows (from AVP analog outs) that it contains real content going up to 24kHz then a sharp brickwall cut. Then immediately above 24kHz there is clearly an reverse reflection which gradually diminishes due to the slow roll-off filter attenuation.

3. Re-sample the 48kHz file (used in 2. above) back up to 96kHz. Spectral playback analysis from AVP analog outs shows same sharp cut at 24kHz and flat transition band up to 48kHz containing no info. Not much imaging thereafter.

Conclusion: The AKM DACs produce reflection 'images' directly above the Nyquist frequency of the flac being played through it.

Note: Whatever type of "oversampling" may be happening inside an AKM DAC, it is clearly not extending the transition band of LPCM signal beyond that of the original PCM file sample rate. The evidence is on the images which show up clearly just after the original Nyquist freq of the Flac file itself.

Ok, many thanks for explaining.

Just a couple of questions:

When you say the 48kHz output “contains real content going up to 24kHz then a sharp brickwall cut” do you mean that the brickwall filter doesn’t begin until 24kHz? That seems very strange.

Or to put it another way: What’s the amplitude of the output signal at Nyquist when playing back at a sample rate of 48kHz?

What’s the amplitude of the images relative to the amplitude of the signal?

And why does the DAC use a slow roll-off filter at 96kHz vs a brick wall filter at 48kHz? Is that something you are configuring yourself or does the AVP select the anti-imaging filter automatically based on the sample rate?

 

[Blumlein 88]

Create a -4 db white noise file. Record that and you'll have a clear picture of the filter response. If you do a 96 khz file, you'll have to record at 192 khz.

You'll get something like this:

Here is one for a DAC with several filter choices.

 

[Martin_320]

Ok, many thanks for explaining.

Just a couple of questions:

When you say the 48kHz output “contains real content going up to 24kHz then a sharp brickwall cut” do you mean that the brickwall filter doesn’t begin until 24kHz? That seems very strange.

Or to put it another way: What’s the amplitude of the output signal at Nyquist when playing back at a sample rate of 48kHz?

What’s the amplitude of the images relative to the amplitude of the signal?

And why does the DAC use a slow roll-off filter at 96kHz vs a brick wall filter at 48kHz? Is that something you are configuring yourself or does the AVP select the anti-imaging filter automatically based on the sample rate?

The steep ("brickwall") attenuation which I refer to is 'burned' into the flac by Studio-One downsampling it from 96kHz to 48kHz.
This cut actually happens just before 24kHz, not actually dead on the Nyquist itself.
Of course, the 'slow' slope attenuation by the DAC happens as well, and starts a bit earlier.

 

[andreasmaaan]

The steep ("brickwall") attenuation which I refer to is 'burned' into the flac by Studio-One downsampling it from 96kHz to 48kHz.
This cut actually happens just before 24kHz, not actually dead on the Nyquist itself.
Of course, the 'slow' slope attenuation by the DAC happens as well, and starts a bit earlier.

I see, thanks. It’s the “slow” anti-imaging filter that’s almost certainly the problem. Does the AVP not allow you to use a proper filter?