advertisement


You pre-rang too soon

adamdea

You are not a sound quality evaluation device
With all the broo ha ha about hi rez, new formats and digital filters over the last 20 odd years, much of the fuss has boiled down it a single issue: does pre ringing matter?

[to skip the rant, just follow the link http://archimago.blogspot.co.uk/2018/01/audiophile-myth-260-detestable-digital.html ]

The main objection which could be levelled against 16/44 was that it requires filtering somewhere around the limits of human hearing. Orthodox theory has it that this should be accomplished by an infinitely steep, infinitely long linear phase filter equal to a mathematical sinc function. This filter "rings" before and after the signal for an infinitely long time, albeit at an infinitely narrow bandwidth. To put it another way, every sampling value affects the value of the reconstructed signal at all times (not just at that instant). [but that is not necessarily a bad thing, see below]

Recently it has been fashionable either to do something very unorthodox (a lazy filter a la MQA) or something heroically over specified (a la chord). The raison d' etre of the former is to avoid pre-ringing which is an inevitable property of the orthodox filter. It has long been hypothesised that this was what was wrong with 16/44 -that it involved adding something which was not there in the original signal pre-sampling, not just that it involved filtering out certain frequencies. The latter approach gets closer to mathematical orthodoxy, and therefore inevitably involves more and more "pre-ringing". Even John Atkinson, in one of his cute footnotes, has acknowledged that it might be thought odd to give rave reviews to both the solution and the amplification of the problem; not that that stopped him.

So is there a problem to cure? if not why are we bothering with solutions?

All along in every article I have ever read about the benefits of NOS dacs, minimum phase, hirez, etc etc, all we get to show the "problem" being "cured" is the impulse response which is the time domain representation of a filter. It represents what the filter would in theory do if faced with a dirac delta function -ie an infinitely short infinitely powerful impulse. It is not a real world signal- the filter and anyone close by would be evaporated by the force of the blast (don't forget that sound is an air-pressure wave). As regards a dac's filter, it is also an illegal signal (the sampling theorem is about band-limited signals).

You cannot therefore judge the effect of a real world signal on a filter in the time domain just by looking at the amount of ringing in its impulse response. Equally you cannot have an accurate reconstruction filter without some ringing in the impulse response. Why? because otherwise you can't fill in the gaps between the sampling instants. No ringing = wrong; and not just wrong in the frequency domain, but also wrong in the time domain.

So the ringing in the impulse response in fact maps to what the filter is doing right (reconstructing what happening between the sampling instants) as well as what it is or might be doing wrong (inventing a signal before or after the sampling instant which was not in fact there in the real signal). But which of these is it in fact doing when presented with a real musical signal? You would thinking that someone in the audio press would have tried to work this out. Unless I have missed something, you would be wrong.

As far as I can tell no one ever tries to show pre-ringing of an anti aliasing or anti imaging filter using a real musical signal. I once asked on Hydrogen audio and someone demonstrated using an acoustic impulse like a spark or a gunshot. Of course pre-ringing is a real problem and is known to those who devise perceptual codecs. But that problem was encountered with filters in the audible range where there is loads of energy in audio signals. Where is the evidence involving a 16/44 filter?

Over on Stereophile Jim Austin is now two parts into his allegedly in depth investigation into MQA and still uses impulse responses to judge filters by what night be called "visual inspection".

Why does no one ever try to show pre-ringing of a 16/44 filter using a real world signal? Let's assume that they are intelligent and inquiring. Well perhaps it's because you would struggle. You see pre-ringing in the impulse response because the assumed signal has infinite bandwidth (at equal amplitude no less) AND complete silence before and after. I and others have been going on about this for years-where is the real world example which shows what's wrong with an orthodox linear phase filter around 22kHz? None of the fancy filters or hi rez for that matter, could be justified unless/save to the extent that there were material pre-ringing to be removed. (of course that's before we even get onto the issue of audibility at the ringing frequency and masking).

Well at last someone has actually done a neat demonstration of pre -ringing (presence and absence) . It's the indomitable Archimago. I think it's a shame that this has not attracted much attention. Perhaps it was the wrong time of year.
http://archimago.blogspot.co.uk/2018/01/audiophile-myth-260-detestable-digital.html
If this is not a first then I would be delighted to see the antecedents.

So, is that it? Does it explain how unorthodox filters (or hi rez) "sound" better on all kinds of material (to some people?)
 
Only scanned though the archmagio page thus far. (Must admit the background/text colour choices her makes do make the content hard for me to read. Wish he'd think of people with imperfect sight.) However what I've picked up thus far I'd agree with.

The problem with clipped signals is that they - if assumed to all be true audio data - lose the real info *and* give the reconstruction system ways to have to 'guess'. Which given the loss is likely to be a wrong guess. Whatever it does, either it has to output HF that isn't within the input LPCM range or waggle the result to avoid the 'sharp edges' which digital-value clipping produced. The real solution is to shoot the idiots who sell such clipped versions in the first place.

One correction to the OP. An impulse isn't 'illegal'. In general there will be an input signal pattern that can cause an ADC to output a single non-zero sample impulse. This comes from matched filter theory. However this is pretty unlikely to happen with real world audio rather than electronically generated waveforms.

The point of the sinc (ringing) filter for a DAC is that it simply passes on the filtering done by the *ADC*. Then up to the people making the recording to get this right. That's what you're paying them for.
 
One correction to the OP. An impulse isn't 'illegal'. In general there will be an input signal pattern that can cause an ADC to output a single non-zero sample impulse. This comes from matched filter theory. However this is pretty unlikely to happen with real world audio rather than electronically generated waveforms.
I'm intrigued by this- I had no idea that this was possible with a signal which is band limited at nyquist.
 
With all the broo ha ha about hi rez, new formats and digital filters over the last 20 odd years, much of the fuss has boiled down it a single issue: does pre ringing matter?
...
So, is that it? Does it explain how unorthodox filters (or hi rez) "sound" better on all kinds of material (to some people?)

Thanks for a really nice "rant" as you put it. People may like the way some DAC filters sound and that's always a valid matter of personal preference.

Nevertheless, as you write, the pre-ringing of a DAC filter in response to an impulse indicates to me that the filter is working properly in a technical sense. And as in the quoted article, I don't know how such an impulse arises with any real-world well-produced input. If you care for the technology (and I am not addressing personal preference here), when you break the basic theorem of digital audio then that's the problem to cure rather than applying a sticking plaster to a symptom of the fundamental problem.

I also recall Keith Howard's 2006 attempt (https://www.stereophile.com/features/106ringing/index.html) to see if people could hear the difference between various filters. The results were not too conclusive.
 
I'm intrigued by this- I had no idea that this was possible with a signal which is band limited at nyquist.

IIUC it's really a matter of how close you can get practice to come close to theory.

As general background: If the impulse response is a perfectly symmetric one with a time range of +/-T the implication is that the group delay in the filter is T at all frequencies. This is nominally so for ADCs or DACs or other resampling processes that have symmetric time responses. You can also argue using time reversal arguments.

The 'matched filter' argument is based on the argument that if you inject a pattern that is the time-reverse of the process's impulse response the output will peak at a (delayed) time determined by the timing of the components - in time or frequency domain. In effect all the levels (samples or frequency contributions) in the time pattern you put in conspire to put all the energy at that output moment.

In practice this may fall down for reasons that I suspect someone like Watts might point out. To approach perfection you need a very wide response pattern and input pattern, with all the values just right.

Think of it this way, though. If you inject a perfect impulse into a ADC with a sinc response, you get a sinc of the relevant spread. If you then poke that into a DAC with the same response you get from an impulse. It can't change any of the frequencies, phases, or amplitudes in what it was given. Then note that a sinc pattern which was time aligned to peak at a sample instant is actually an impulse set. :)

In reality, the shapes and values aren't perfect. But since a sinc function is time symmetric, the above does, indeed, use a 'time reversed' input to do this.

That's the argument anyway. Although I admit I've never tried to do it.
 
Jim, Stereophile have a forum, it's here:
https://www.stereophile.com/forum

If you think Jim Austin is talking trash, why not raise it on their forum and attempt to have the debate directly, then simply let us have a link to how it goes.
Surely that's not only more polite to Jim, as it gives him a much better opportunity to respond, but also might lead to a faster clarification.
 
I also recall Keith Howard's 2006 attempt (https://www.stereophile.com/features/106ringing/index.html) to see if people could hear the difference between various filters. The results were not too conclusive.
Yes and not just any old people. Strangely nowadays when they know what the "correct" answer is they seem more certain.
You may also be aware of Werner's distributed file filter test on this forum, and Archimago's minimum phase vs linear phase test. These also produced results which indicate that there is no systematic preference for minimum phase.
 
Jim, Stereophile have a forum, it's here:
https://www.stereophile.com/forum

If you think Jim Austin is talking trash, why not raise it on their forum and attempt to have the debate directly, then simply let us have a link to how it goes.
Surely that's not only more polite to Jim, as it gives him a much better opportunity to respond, but also might lead to a faster clarification.
Why not have a look yourself. You might find that the points had been made and ignored.
https://www.stereophile.com/content/mqa-tested-part-2-fold
 
Jim, Stereophile have a forum, it's here:
https://www.stereophile.com/forum

If you think Jim Austin is talking trash, why not raise it on their forum and attempt to have the debate directly, then simply let us have a link to how it goes.
Surely that's not only more polite to Jim, as it gives him a much better opportunity to respond, but also might lead to a faster clarification.

Afraid I don't know who Jim Austin is, or why you think I "think" he is "talking trash". So far as I am concerned, using impulse functions to examine a process may be a reasonable thing to do. My earlier comments were to clarify a point to Adam
 
I've now had a scan though the Stereophile thread on their forum. Not done so in detail - but I've not yet read Archamio's page either properly as yet. However

Adamea: I noticed you gave a URL for one of my pages prompted by MQA. It might be useful if you also pointed out the related ones as some people may not see them otherwise

http://www.audiomisc.co.uk/MQA/intoshape/NoiseShapingHighRez.html

http://www.audiomisc.co.uk/MQA/cool/bitfreezing.html

http://www.audiomisc.co.uk/MQA/bits/Stacking.html

If someone wants to get a full idea of my views on MQA, they'd need to read the above as well. But my earlier comments here were directed at a more basic level wrt the nature of impulse functions and responses as set by one of the axioms of information theory, not MQA. Nor, indeed, Jim Austin.
 
Adamea: I noticed you gave a URL for one of my pages prompted by MQA. It might be useful if you also pointed out the related ones as some people may not see them otherwise

http://www.audiomisc.co.uk/MQA/intoshape/NoiseShapingHighRez.html

http://www.audiomisc.co.uk/MQA/cool/bitfreezing.html

http://www.audiomisc.co.uk/MQA/bits/Stacking.html

If someone wants to get a full idea of my views on MQA, they'd need to read the above as well. But my earlier comments here were directed at a more basic level wrt the nature of impulse functions and responses as set by one of the axioms of information theory, not MQA. Nor, indeed, Jim Austin.
Done.
 
Thanks for pointing this out. It's good to see some real world analysis of this, it's also helped me improve my limited understanding a bit.
 
1) for a DAC the Sinc function is the ideal reconstruction filter. The sampling theorem says so. (If you don't believe it read it and understand it.) It is the only filter that guarantees a full and correct, in frequency and in time, rendering of the digitally-stored samples to the analogue domain.
The only requirement is that, at recording time, the signal is band-limited prior to sampling. In other words: there must be zero content at Fs/2 and above.

2) A consequence of the above is that the onus is entirely placed on the ADC anti-aliasing filter, or on the mastering downsampler. It must guarantee a sufficient suppression at Fs/2 and above. So it demands a steep filter, which is ringy.
On the other hand, if the recording filter does its job, then the DAC filter can be allowed to be less than ideal.

3) Such filters ring at their transition frequency. For CD rate and above these frequencies are totally inaudible to the utter majority of healthy adults. If you want to run critical listening tests, use healthy, trained teenagers. Do not rely on balding fifty-somethings with a beer belly and a forum account. Or a press card.

4) There is a relation between the audibility of ringing and a filter's transition bandwidth. This connects directly to the inherent timeconstant of the cochlea at said frequency. Once the transition band is wide enough the ringing becomes inaudible, even when the ringing frequency is smack in the middle of the audible band (e.g. 3kHz). We are speaking here of transition widths of 100s of Hz in the midrange, and a few kHz in the top octave. Hardly an impediment.

5) Most music, even transients, does not contain a lot of energy at 20kHz, compared to what lives below 3kHz or so.
If you look at a drum beat of whatever, you'll see that it really is not all-nothing and then a sudden rise to maximum. Even a so-called transient generally starts at a feeble level and then swings in to a maximum in a couple of cycles. There is really not much that can provoke visible ringing in a waveform post-sampling.
(In fact the only time I have seen it with a real-world input was with ticks on an LP.)
 
4) There is a relation between the audibility of ringing and a filter's transition bandwidth. This connects directly to the inherent timeconstant of the cochlea at said frequency. Once the transition band is wide enough the ringing becomes inaudible, even when the ringing frequency is smack in the middle of the audible band (e.g. 3kHz). We are speaking here of transition widths of 100s of Hz in the midrange, and a few kHz in the top octave. Hardly an impediment.
All very interesting. I can remember you mentioning this point in the past but I've not quite got my mind absolutely round it. Do you have any references for this? Is it connected with temporal summation in a neuron?
 
Is it connected with temporal summation in a neuron?

No. More the temporal properties of the equivalent band pass filters of the basilar membrane. JJ posted some about this, and probably has it in more detail in one of his published presentations. Similar considerations must pop up in early research into filter banks for perceptual coding, e.g. MP3. But that brings you back to JJ and Brandenburg.
 
No. More the temporal properties of the equivalent band pass filters of the basilar membrane. JJ posted some about this, and probably has it in more detail in one of his published presentations. Similar considerations must pop up in early research into filter banks for perceptual coding, e.g. MP3. But that brings you back to JJ and Brandenburg.
I have some of his presentations somewhere
I spose I could have a go asking on HA, but it can be very tiring trying to get information there.
I wonder whether its possible to establish the time constant of the 20 Khz + band (!)
 


advertisement


Back
Top