Digital filtering

[darrenyeats]

New thread jumping off from my OT tangent in the latter part of https://pinkfishmedia.net/forum/threads/noise-shaping.198251/. For those interested (both of you!) see posts there from me and Jim for background.

Now using intermediate phase (like "-p 25" in SoX)! That's means half way between minimum and linear phase in terms of phase distortion.

Before I was treating linear phase as good/home base. After more listening, I've now decided minimum phase is good/home base for me. Extrapolating from Archimago's blog, "99% of minimum phase goodness" for half the phase distortion.

Will report back long-term.

I'm still happy with intermediate phase (-p 25) after many weeks. Below is the difference between minimum phase (-p 0) and intermediate phase (like -p 25) in SoX. My thoughts about it below.
1. Notice how intermediate phase doesn't have any visible pre-ringing. With SoX it seems very sensible to use intermediate phase instead of minimum phase.
2. Some "minimum phase filters" provided in DACs and SRC apps may really be "minimum phase-type filters" i.e. intermediate phase (and indeed ISTR that phrase in a review) rather than technically minimum phase.
3. Both the below are with wide-bandwidth very steep filters near 22kHz. I like a flat-to-19kHz filter which is less steep, so the ringing would be somewhat less in both cases.

I've rigged up my hacked LMS SoX-side volume control to switch seamlessly between filters, and I'm homing in on the best tracks on which to do some blind testing of linear versus intermediate phase filter (I have only 1 or 2 possibles so far). The quick A/B has revealed the difference generally isn't night and day: I need a gross, quickly heard difference to have a chance with blind testing (I find it long and very unpleasant, and my brain shuts down quickly!)

 

[John Phillips]

For information (but relevant to my kit, below), the AKM AK4497EQ DAC chip has six filter settings. See the "Sound Color Digital Filter (SCDF)" table circa 2/3 way down this page (look at the impulse responses but ignore the highly fanciful marketing copy).

• The top three impulse responses (1, 2 & 3) have wide-band frequency filters.
  
• The bottom three impulse responses (4, 5 & 6) have narrow-band frequency (probably proper Nyquist) filters
  
• Impulse responses 3 and 6 seem to be linear phase filters (symmetrical filter coefficients with the corresponding high group delay)
  
• Impulse responses 2 and 5 seem to be minimum phase filters (asymmetrical filter coefficients with very small group delay)
  
• Impulse responses 4 seems to be an intermediate phase filter (asymmetrical filter coefficients with an intermediate group delay)
  
• Impulse response 1 - I have no idea how to describe that filter.

My ATC CDA2 Mk2 uses an AKM AK4490EQ DAC fixed on "short delay sharp roll-off". AKA minimum phase - Impulse response 5. And just a note that this DAC chip only has five of the six responses above - it lacks intermediate phase.

On streaming I drive the CDA2's USB input with a R.Pi running picoreplayer, which has SoX. I have done a few very badly controlled tests using SoX's oversampling and filter phase control to "override" the CDA2's fixed filter. There are no "night and day" differences that I perceive.

EDIT: typo (DAC name).

 

[John Phillips]

FYI, Archimago's recent musings on filtering are here. See "III. Let's go "Extremus" with the SoX digital filter!"

 

[darrenyeats]

Thanks John, I've now left some comments there.

 

[John Phillips]

Even though I think @darrenyeats tastes may lie in another direction, there's an interesting Keith Howard article in a recent issue of HiFiCritic about Chord's M-Scaler and Rob Watts' approach to digital filtering.

I find the article very enlightening, but it does not really answer the questions it raises about just how "heroic" you need to be in digital filter length for there to be no more audible differences, never mind dealing with personal preferences. I listened to the M-scaler demo at Audio Show East this year and didn't find it a convincing differentiator. But that's possibly typical of a show demo.

Rob Watts for Chord is definitely at the opposite end of the spectrum to Bob Stuart for Meridian when it comes to digital filtering.

 

[darrenyeats]

I was at the same demo. Although I didn't find the DAVE with or without M Scaler especially musical, I did hear a difference! Mainly sound stage got (a lot) better to my perception. When I sat between the speakers I heard a significant widening with the M Scaler. Later, when I sat in front of the left speaker, I heard the sound stage extending toward and even behind me! To me this was annoying on several levels! I don't blame the ATC 40s for lack of musicality though ...

Especially since the active ATC 100 in other room were very musical and good in every way except ... I found the top end on most of the digital recordings was, well, digital. No complaints on vinyl though, loved it! But as you wrote, the ATC CDA2 Mk2 front end uses minimum phase-style filter, so it might seem I'm contradicting myself here. Except I'm not because I'm saying for me I want a minimum phase-style filter but I don't mean it's enough on its own! It just helps. Ringing from a steep filter isn't good, it's just not as bad as pre-ringing from a steep filter (or, HF imaging).

On that note, I feel hi res is superior all else being equal - but other factors still come into play. Recording quality, acoustics, speakers, amps, front end, sampling rate, filter phase type all have their part, some are more often a limiting factor than others. It gets complicated because certain DACs might work better with certain sampling rates and so on, so there's an interaction between factors. I know with my DAC, for example, that if I up-sample on my server using SoX -v (very high quality) option, the result is better than letting the DAC up-sample internally. So one might blame CD format for a certain sound, but it sounds different when processed differently (we've come full circle to the M Scaler!)

But I'd love to hear the same Chord DAVE and M Scaler demo with minimum phase-style filter, for kicks.

 

[Jim Audiomisc]

Even though I think @darrenyeats tastes may lie in another direction, there's an interesting Keith Howard article in a recent issue of HiFiCritic about Chord's M-Scaler and Rob Watts' approach to digital filtering.

I find the article very enlightening, but it does not really answer the questions it raises about just how "heroic" you need to be in digital filter length for there to be no more audible differences, never mind dealing with personal preferences. I listened to the M-scaler demo at Audio Show East this year and didn't find it a convincing differentiator. But that's possibly typical of a show demo.

Rob Watts for Chord is definitely at the opposite end of the spectrum to Bob Stuart for Meridian when it comes to digital filtering.

In reality, the question ends up depending on the filter function of the ADC. Once the DAC filter is distinctly longer the ADC will determine the bulk of any alterations due to 'digital filter' behaviours.

In that limited sense Stuart's 'MQA' argument is in principle correct wrt 'correction'. But duly falls down when you realise the impulse response of microphones, etc, are vastly more severe. Hence the basic real-world engineering point that it becomes irrelevant to try and make one link in a chain 'absolutely perfect' when other links have more gross problems you're ignoring.

 

[darrenyeats]

I agree in that there's no magic bullet - this is perhaps the tl;dr version of my last post.

If you can point me to some references re: microphones that would be great. On the tweeter end, ISTM many tweeters ring less than a non-leaky 44.1kHz filter and have a decent phase response.

 

[Jim Audiomisc]

Alas, reliable info on most commonly used microphones is scarce. I've been trying to find more than a tiny handful of examples for ages! You can find data on things like 'measurement' microphones of the kind someone in a lab or R3 might use at times. But for most studio mics things get harder. One person in this area I talked to pointed out that at least one mic company he knew of 'cut and pasted' the results from one model's spec to another. None seem to give a reliable time domain measurement. But you can often see a lift and then fall well *below* 20kHz in published results, which then cut off at or *below* 20kHz in many cases for classic mics. No sign of the effects of distance or angle of incidence, either, in most cases.

Bascially, many studio mics are sold on 'sound'. Fine in itself, but it means a lack of reliable info unless we simply sigh and assume that the mic is a part of the sound of a piano or violin!

Speakers are another issue, of course. There the main effects may be directional, room dependent, and due to things like box size/shape and the spacing of the speaker units along with the crossover. So affect lower frequencies on a much longer time scale. In the end, this gets solved by the user deciding what suits them best.

 

[John Phillips]

In reality, the question ends up depending on the filter function of the ADC. Once the DAC filter is distinctly longer the ADC will determine the bulk of any alterations due to 'digital filter' behaviours.

In that limited sense Stuart's 'MQA' argument is in principle correct wrt 'correction'. But duly falls down when you realise the impulse response of microphones, etc, are vastly more severe. Hence the basic real-world engineering point that it becomes irrelevant to try and make one link in a chain 'absolutely perfect' when other links have more gross problems you're ignoring.

I think it comes down to where in the chain you decide to put a "reference point".

An argument I find reasonably persuasive is that whenever the final mastering engineer says "go for production" then that is "the art". Everything beforehand, whether the performance itself, or technical matters like the impulse responses of the microphones used, the responses of the Nyquist filters in front of the ADCs, or the digital filtering/equalization applied in the mastering suite, has at that point, for better or worse, become a part of "producing the art" according to the skills of the people involved.

The rest of the chain is for me to decide on how I want to "reproduce the art". That may include making changes under my control to enhance how I want to enjoy the art, such as (deprecated these days) tone / balance controls or other matters such as using favourite digital filter designs whatever they are.

But I entirely agree that looking to achieve heroic degrees of technical perfection in the reproduction chain (according to some standard for technical perfection) can go far too far considering the uncorrectable technical imperfections not under my control.

 

[John Phillips]

But I'd love to hear the same Chord DAVE and M Scaler demo with minimum phase-style filter, for kicks.

The Chord M-scaler does have a low latency video mode, I think. To avoid audio/video lip-sync problems from the group delay of a 1 million tap linear phase filter.

I am not sure if that is achieved through just a shorter linear phase filter or through a minimum phase filter. And I am not sure how that would interact with the 164,000-tap linear phase filter in the Chord Dave.

One could ask at another demonstration.

 

[darrenyeats]

Hi John,
I found a manual here https://chordelectronics.co.uk/wp-content/uploads/2018/07/M-Scaler-Manual.pdf - the number of taps is reduced in video mode, so looks like a linear filter still.

 

[Jim Audiomisc]

I think it comes down to where in the chain you decide to put a "reference point".

Agreed. But for me this is "what I heard myself when at the concert hall during a live performance". i.e. the reference R3 tend to use. Hence factors like the mics certainly are in the chain.

In situations where the producer has 'confected' a sound *when listening to his listening room monitors* and you have no way to tell what he heard or wanted, then changing one decent DAC filter for another is essentially what I've described in the past: a form of 'tone control'.

Personally, I am quite happy to have tone controls on the systems I use. And I do, indeed, alter them to suit a given bit of music and my taste/judgement. So this isn't an implied criticism on my part. Just what seems to me to be a realistic viewpoint. Yes, if someone prefers one DAC filter to another, then that's the one they should choose.

 

[darrenyeats]

Jim, good point.

Even using John's definition my reference point would be the final mix. The mastering is often done by different organisations and sometimes much later and I don't consider it the art.

If the final mix was tape or high res then I consider it my business to correct or compensate for what I perceive as shenanigans after that, e.g. mastering down to 44.1kHz.

 

[Jim Audiomisc]

I think it a pity that tone controls have become regarded as 'verboten' over the years. Well designed and implimented types can be very handy. They also give scope for tweaks that the usual 'suspects' in terms of DAC filters do not. For example the old QUAD bass lift/shelf on a 34, etc. And in principle very precise/subtle/flexible ones can be done using DSP.

 

[John Phillips]

I was thinking about Chord’s approach to digital filtering in a DAC.

Without knowledge of the ADC, the sampling theorem demands that a perfect DAC reconstruction filter have a “brick-wall” characteristic. AIUI the only mathematically correct way to do this is to have a perfect “sinc” filter.

But you cannot implement a perfect sinc filter. One reason is an infinite delay time.

So, you have to abandon perfection and look for “good enough”. For the sampling theorem: enough rejection of aliases; For the audio: flat enough frequency response and sufficiently constant delay (or perhaps something else according to taste).

"Good enough" means you no longer need to use a sinc filter. Many types of filter can be designed to approach the requirements closely enough.

Chord pursues the sinc filter to increasing accuracy and makes it a marketing point. And I like the way their DACs are very free from the obvious engineering defects. But the filter coefficients fall off slowly as Keith Howard shows in the HiFi Critic article above. That makes me wonder if it’s really better engineering to select a different filter to achieve “good enough” with less hardware.

 

[Jim Audiomisc]

Agree wrt needing to know the ADC details. However the sinc doesn't need to be 'infinite' because all real recordings have a finite duration. And will have been done with an ADC that itself doesn't reach 'perfection'. So you only need a DAC filter that is 'good enough' not to do harm that rises above being swamped by the recording limitations.

In practice, the choice of ADC (and all kinds of other factors that alter what ends up being given to you to play) means that the most sensible options tend to be:

1) Choose a given 'minimum harm' ADC filter like the sinc-approach (or whatever you prefer given the *rest* of your replay kit altering things anyway).

or

2) Be able to change the DAC filter case-by-case as you fancy. 'Tone controls' approach.

From my initial comments here you can in principle upsample the entire recording using one big FFT per channel and match the length of the 'reconstruction filter' to the length of the given recording. Just that you need a setup that can do the FFT and store the data until you can play it. (The 'sinc' result follows from FT arguments.)

 

[darrenyeats]

The problem here is we're talking about perfect implementation of "something". But for 16/44 unfortunately that something is a "bad something" i.e. a steep filter for 22kHz! When music has energy present in such a transition band, this leads inevitably to ringing either before or after, or imaging, or some combination.

Personally I find the Chord DACs unmusical, but OTOH I did experience increasing number of taps to improve the sound-stage.

I would dearly love to hear the DAVE+M Scaler but with a minimum-phase style filter. I wonder if the sound-stage would collapse or not. I assume Mr Watts has tried it and didn't like it ...?

I suspect the only way to fix this tension is with hi res source. For me, the least bad camp is the causal AKA minimum phase-style AKA analogue-like filter at some point in the engineering chain, especially for 16/44. Since the only bit I control is playback I use it there, although I accept it gives sub-optimal results in some cases e.g. when the engineers already employed it in part or whole!

It's worth mentioning again that filter phase doesn't change the frequency response, and for this reason I don't consider it a tone control in the normal sense.

 

[Jim Audiomisc]

The problem here is we're talking about perfect implementation of "something". But for 16/44 unfortunately that something is a "bad something" i.e. a steep filter with a transition band starting somewhere not too far from 20kHz and ending 22kHz! When music has energy present in this band, this leads inevitably to ringing either before or after, or imaging, or some combination.

It's worth mentioning again that filter phase doesn't change the frequency response, and for this reason I don't consider it a tone control in the normal sense.

A difficulty here is the curiously absent nature of much reliable data on the performance of microphones. However those few I *have* seen measurement results on show bandwidths somewhat below 22kHz and hence may 'ring' at lower frequencies anyway. That in turn assumes that much of the sound hitting them contains significant HF at or above 22kHz and realising that the *sources* are causal, so may time-smear the HF anyway.

The phase effects of the filters will alter the time alignment. It then becomes a matter of debate/circumstance/etc if it may or may not matter. Time smear at HF might be good for replay as it can reduce peak/mean ratios that amps and speakers are given.

FWIW personally I suspect such time smearing matters more at the LF end of the spectrum as there a given phase shift corresponds to a much longer time shift. But in the end I stick with feeling this is largely a matter of what each person/system finds best. That's why I think it a good idea for people to experiment and decide.