"I can hear them church bells pre-ringing!" - A (mostly) time domain investigation into DAC-like reconstruction filters

[antcollinet]

Here are two tests/experiments/show cases, with synthetic signals and real DAC. Seems to fit the topic, so maybe someone will find that interesting.

Below I'll be using FiiO K3 ESS, so to start with, here's its impulse response. It uses a linear phase filter.

View attachment 517285

I wanted to check (or show) that the linear filter doesn't introduce (pre)ringing if there is no content at the cutoff frequency. I generated this 20k-bandlimited pulse at 764k sampling frequency and converted it to 44.1k with SoX rate 44.1k (default settings, so also linear phase):

View attachment 517286

Then I played this 44.1k version on FiiO K3 and recorded on E1DA ADCiso again at 764k sampling frequency. Here is the original at the top and the recording at the bottom:

View attachment 517287

and zoomed in to 0.1 (-20 dBFS):

View attachment 517288

and zoomed in to 0.001 (-60 dBFS):

View attachment 517289

and this, AFAICT, is the shaped noise produced by the DAC. So no pre-ringing in sight.

---

Next, I wanted to check what level of pre-ringing can be expected from a wide band transient. For this I wanted to use something that more closely resembles a transient that could actually exist. Not a single sample. I generated this pulse at 764k sampling frequency and converted it to 44.1k like earlier:

View attachment 517290

My criterium for "more realistic" was this falling of the level as the frequency increases. To be honest I don't know if that is actually anything close to realistic or not . Anyway, played it on FiiO, recorded on E1DA, the original is again at the top and the recording at the bottom:

View attachment 517291

and zoomed in to 0.1 (-20 dBFS):

View attachment 517292

We get pre-ringing below 0.01 (-40 dBFS) level. 4 periods in 190 µs gives as 21052 Hz. Probably not-so-coincidently, that's a -6 dB point in SoX's rate frequency response:

View attachment 517293

I also cut 2 ms before the pulse and did FFT. The resolution is not so good, but with a longer cut it starts to drown in the noise:

View attachment 517294

Nice work - and another illustration that ringing is not "created" by the filter. Those frequencies are there in the pre-filtered signal - it is just they are revealed when high frequencies (that "cancel" them) are removed by the filter. A subtle but meaningful difference when we consider that the ear is also a low pass filter.

 

[Csaba Kelemen]

We need to prove that the deviation of the reconstructed signal from the original is guaranteed to be below a threshold at all times. I would only call a DAC transparent if it guarantees that this threshold is below the LSB. I don't see the proof in your explanation.

 

[voodooless]

I would only call a DAC transparent if it guarantees that this threshold is below the LSB. I don't see the proof in your explanation.

That's an issue of definition. "Transparent" for most of us means: audibly transparent.

 

[antcollinet]

We need to prove that the deviation of the reconstructed signal from the original is guaranteed to be below a threshold at all times. I would only call a DAC transparent if it guarantees that this threshold is below the LSB. I don't see the proof in your explanation.

Not when the LSB is lost 20dB below the analogue noise floor of the DAC, and is at least 30dB lower than any human ear can detect.

Transparency is about audibility - not LSBs. Otherwise you could build an 8 bit DAC and call it transparent because it is perfect to an LSB - even though the quantisation noise would be very audible.

EDIT : also you need to quote who you are replying to - it is far from obvious. You can hit "reply" bottom left of the post you are referring to - or select a section of text in the post - hit the reply that pops up next to that - and only that text will be quoted rather than the whole post.

 

[Csaba Kelemen]

That's an issue of definition. "Transparent" for most of us means: audibly transparent.

I expect the definition of precision accepted in technical and scientific life, not arbitrary "audible" rhetoric.

 

[Csaba Kelemen]

Not when the LSB is lost 20dB below the analogue noise floor of the DAC, and is at least 30dB lower than any human ear can detect.

Transparency is about audibility - not LSBs. Otherwise you could build an 8 bit DAC and call it transparent because it is perfect to an LSB - even though the quantisation noise would be very audible.

EDIT : also you need to quote who you are replying to - it is far from obvious. You can hit "reply" bottom left of the post you are referring to - or select a section of text in the post - hit the reply that pops up next to that - and only that text will be quoted rather than the whole post.

Why do you reject the concept of a transparent 8-bit DAC, a transparent 16-bit DAC, or a transparent 24-bit DAC? Would that be too exact for you? Isn't it just that it requires mathematical training? Isn't it just that you can't tell the audiophile what he can and can't hear. Isn't it just that real scientific work needs to be done instead of "gaslighting"?

 

[voodooless]

arbitrary "audible" rhetoric.

It's not arbitrary. Audibility thresholds are pretty well established.

 

[antcollinet]

Why do you reject the concept of a transparent 8-bit DAC, a transparent 16-bit DAC, or a transparent 24-bit DAC? Would that be too exact for you? Isn't it just that it requires mathematical training? Isn't it just that you can't tell the audiophile what he can and can't hear. Isn't it just that real scientific work needs to be done instead of "gaslighting"?

Transparency to most people here is an audibility topic. Whether or not what a human with good hearing, can hear an audibly perfect representation of the music: Whether what comes out of a device is audibly identical to what goes into it.

Here is our definition of audibility thresholds. Human hearing has neither infinite sensitivity, nor infinite bandwidth. When you look into it is is surprising how inaccurate it is in some areas - particularly our ability to distinguish separate tones when close together:

Audibility thresholds of amp and DAC measurements

Introduction A recurring theme on ASR is whether or not the various measured qualities of the devices are audible. In this post, I'll present some clear and visual thresholds for when those imperfections can be considered a potentially audible concern. I will not explain the basics of amp/DAC...

www.audiosciencereview.com

Being better than that is meaningless - we can't hear it - so it doesn't matter. Being audibly imperfect can never be transparency - no matter how compliant with device specifications. What comes out is changed compared with what goes in.


Feel free to have your own definition, but it is pretty pointless having your own unique definition of any word.

 

[Csaba Kelemen]

It's not arbitrary. Audibility thresholds are pretty well established.

Okay. Where is the proof that the ESS reconstruction filters are inaudible and accurate (the error is below 120db). I've been looking for this for a long time and I doubt that such a thing exists. What I find, however, is the "gaslighting" of the supporters of big tap numbers.

 

[antcollinet]

"gaslighting"

You keep using that word. I don't think it means what you seem to think it does.

 

[Csaba Kelemen]

Transparency to most people here is an audibility topic. Whether or not what a human with good hearing, can hear an audibly perfect representation of the music: Whether what comes out of a device is audibly identical to what goes into it.

Here is our definition of audibility thresholds. Human hearing has neither infinite sensitivity, nor infinite bandwidth. When you look into it is is surprising how inaccurate it is in some areas - particularly our ability to distinguish separate tones when close together:

Audibility thresholds of amp and DAC measurements

Introduction A recurring theme on ASR is whether or not the various measured qualities of the devices are audible. In this post, I'll present some clear and visual thresholds for when those imperfections can be considered a potentially audible concern. I will not explain the basics of amp/DAC...

www.audiosciencereview.com

Being better than that is meaningless - we can't hear it - so it doesn't matter. Being audibly imperfect can never be transparency - no matter how compliant with device specifications. What comes out is changed compared with what goes in.


Feel free to have your own definition, but it is pretty pointless having your own unique definition of any word.

In technical and scientific life, accuracy is important to us, which is almost always characterized by the largest possible deviation from the ideal. So for us, the meaning of 8-bit DAC, 16-bit DAC, etc. is clear. The typical ASR commentator rejects this because he does not have the appropriate scientific skills. He does not know how to design a reconstruction filter with the expected accuracy. How to prove the correctness of the principle.

 

[MaxwellsEq]

typical ASR commentator

Avoid generalizations.

 

[voodooless]

Where is the proof that the ESS reconstruction filters are inaudible and accurate (the error is below 120db).

As you might know, you can't prove this negative. We can only prove a positive diffrence in audibility tests.

What I find, however, is the "gaslighting" of the supporters of big tap numbers.

Ah, so you can prove that they, in fact, make an audible diffrence?

 

[RandomEar]

@Csaba Kelemen, this thread is about ringing in reconstruction filters, not about maximum deviations. I published the filter coefficients so if you would like to test the maximum deviation yourself, you've got all that you need to do this.

 

[voodooless]

I would only call a DAC transparent if it guarantees that this threshold is below the LSB. I don't see the proof in your explanation.

Even if we ignore any audible transparency, no audio DAC in existence has 1 LSB accuracy. The noise floor is at least 10 dB higher, and the LSB usually contains dither anyway.

We need to prove that the deviation of the reconstructed signal from the original is guaranteed to be below a threshold at all times

"at all times"? What exactly does that even mean?

 

[antcollinet]

Ah, so you can prove that they, in fact, make an audible diffrence?

I'm always fascinated by the emergence of the sleepers. He's been here for two years - what has just triggered this stepping out into the light of day with all the usual nonsense.

EDIT - looking at his post history starting a few days ago - I'm going to assume he's purchased the Shiit Magni Mesh, and is now disgruntled that it didn't get the glowing review he believes is deserved. It's OK @Csaba Kelemen it might be "distinctly average" but it is still more than good enough to be transparent to everyone in real world listening.

 

[Hayabusa]

Even if we ignore any audible transparency, no audio DAC in existence has 1 LSB accuracy. The noise floor is at least 10 dB higher, and the LSB usually contains dither anyway.


"at all times"? What exactly does that even mean?

during the whole signal?

 

[Csaba Kelemen]

I'm always fascinated by the emergence of the sleepers.

You cannot distinguish between scientific knowledge and mere speculation, or even ad hominem reasoning. I have the impression that you are deliberately misunderstanding.

 

[Mort]

You cannot distinguish between scientific knowledge and mere speculation, or even ad homminem reasoning. I have the impression that you are deliberately misunderstanding.

And I have the impression you are out of arguments. Care to continue?

 

[Csaba Kelemen]

@Csaba Kelemen, this thread is about ringing in reconstruction filters, not about maximum deviations. I published the filter coefficients so if you would like to test the maximum deviation yourself, you've got all that you need to do this.

The mega tap gurus (Rob Watt -Chord, dCS, HQPlayer, Mike Moffat) argue that supercomputing, or rather huge convolutions, are required for reasonably accurate signal reconstruction. But the typical opinion here is to reject this. This forum also claims scientific foundations in its name. In my opinion, your study did not clarify this issue, it only maintains beliefs. With an argument based on maximum deviation, one could reasonably formulate a valuable criticism. That is why I brought it up.