What Makes Good Filtering?

[Miska]

But this is a sneaky one. So yes I understand. But how do we test this for real merits ? a microphone recording of a test signal ?

 

You take the recording equipment and record Dirac pulses with it, just from the line-in. I can generate the test signals easily using AWG.

 

But for ADC chips all you need to do is take the datasheet and look at the digital filter response plots and specs they give for the digital filter and you already know pretty well how the output is going to look like...

 

For software conversion used at mastering stage (from 96/24 or similar masters), you can just go to the SRC comparison page and see the figures for various different pieces of software. There's also the impulse response plots for Dirac pulses.

[Fokus]

Of course you could sometime try how a maximum-phase FIR would sound like where all the ringing is pre-ringing...

 

Been there done that. Max.Phase against Min.Phase against 96kHz source. Listening panel of ten audiophiles, all using their own systems, some using august systems. Leisurely. Blind.

[Jud]

Been there done that. Max.Phase against Min.Phase against 96kHz source. Listening panel of ten audiophiles, all using their own systems, some using august systems. Leisurely. Blind.

 

Leisurely, blind, and I'm supposing not better than chance identification of differences? The very odd thing in this is that not only do I agree with Mani that (sighted) I believed I could clearly hear differences between HQPlayer filters, but that it surprised the hell out of me how much of a difference I thought there was. Placebo OK, but I don't expect placebo to be surprising!

[Jud]

By the way - having said what I said about placebo, surprises, etc., and having brought up the issue of audibility (necessary I think in a discussion of what makes a good filter), I don't want to get distracted into arguments about this. I am enjoying the "signal to noise ratio" we're getting in this thread and would ask everyone participating to keep it that way. (Note: Anyone here is still welcome to tell me I am crazy for thinking I hear various differences.) By the way, Fokus, something in your last couple of comments has led me to another question: Why is it that you like linear phase? Easier, something about it more "artistically" appealing, or some sort of bad (audible?) effects (e.g., group delay)?

[Superdad]

In reference to the Max Filter Length setting in Advanced iZotope parameters, Jussi wrote:

 

That's what I was trying to say too, it's a maximum value allowed...

 

But the conversion ratio will have it's say on the total equation too, so you have to take that into account when setting the value. I think it can be safely left at default but I'm not sure why the setting is even necessary there.

 

Well I can tell you beyond a doubt that, within Audirvana 1.5.10, the default setting of 2,000,000 sounds quite different than my preferred setting of 1,300,000, and that I can hear the differences down to a granularity of about 75,000. Admittedly this parameter has a vastly lesser effect on SQ than Steepness ("order"), Cut-off, or Pre-ring, which I have optimized (but need to revist) for my DAC at 7 "dB", 1.02, and 0.86 respectively. And I have requested that Alexey Lukin provide finer control of all three as each is still coarser than what I can resolve with my ears and test tracks.

 

I know Peter didn't want this to turn into an iZotope thread, but I did want to point out that the Length parameter does have an effect--different in character than the other controls. Alexey is not here to explain what exactly it does, so I do not think it is fair to dismiss that control as redundant or not necessary.

[Superdad]

And my not-so-subjective opinion : good bye with that because post ringing runs into the next sample (data) and destroys that. Technically it just works like that. So if you think about the ms value of an xxx tap filter and know that after 22us (native 44100) a next sample comes around, you can see that this is not much good. Thus, the post ringing is not so much about echos as always told, but how the next sample is involved. what ? 100s of samples (or 50 according the 117 tap or whatever it was).

 

Peter,

Months ago when you first published on your site the measurment graphs showing that, I got very excited about it because I had never seen someone show or consider that (the post ringing running into the next sample).

[Many of your measurements on that particular page were real eye-openers, and in the right hands could lay waste to so many naysayers. I think you maybe are too modest about some of the things you have discovered and proved there.]

 

So is not that a good argument in favor of shorter (or at least no longer than a certain length/taps) filters? Can you put some numbers to that for us?

 

Regards,

Alex C.

[PeterSt]

Can you put some numbers to that for us?

 

Hey Alex,

 

That would be a too "static" approach IMO. But let me give you a hint about the why :

 

As I pointed out earlier (or tried to), much is about "masking". But this is also about "masking what exactly" and this is quite impossible without looking at the result of the whole, plus it needs a representation of what music would be able "to do" in this case. "To do" = destroy, influence, impact. And now it is all related to the further filter settings itself. I'll try to give an example :

 

You asked about a.o. the number of poles. Well, if you'd see what can happen with changing that to for example the phase, you would also see how rough such a parameter is. Chance that all is right is close to zero. However, the cutoff frequency itself is very granular so it is quite easy to tune the filter with that (per Hz as it is now, but can be 0.1Hz etc.) and see the (linearity of the) phase get right. Meanwhile the ringing can get too long (like dialing a too low frequency and again a few Hz can be enough to destroy) and all is destroyed. Point is : without notice.

So what it needs is a graphical representation and it is quite beyond me (at least at this time) to create that out of a cascade of several filters. Also, in the end it needs the D/A chip (or better : the converter with its (slew rate) speed as a whole) to observe the real net effect of all, so ... undoable for that alone.

And thus, while I headed for a nice set of dials (say like in iZotope) to construct a filter sequence how you want it, it all can not work in practice because you need to see the net result. And you can't.

 

I can because I look through the analyzer so all I can do is make presets, no matter theoretically you as a user could do all. And then still all I can do is look at the output of a DAC which doesn't add anything and is fast enough for all sorts of filters, which is not your DAC.

So it's quite a strange problem, once we see what the real merit of this all could be. It is almost too much of teasing ...

 

Regards,

Peter

[barrows]

The problem with such analysis is that, when considering 16/44.1, any approach to filtering is going to be compromised. That is, there is no "perfect" filter, there are only trade offs. So then the filter developer has to decide which parameters are more important for sound quality.

So, now we see that we are in the realm of purely subjective analysis: is reduction of pre-ringing more important? is extended frequency response more important? is reduction of overall ringing more important? is complete suppression of alias products more important? Who really knows? What we need is more study into psychoacoustics, where, maybe, some of these questions can be answered. But it will not be easy!

Then add the fact that everyone's "reference" system is different, such that some systems may sound better with a minimum phase filter with no pre-ringing, and other systems may sound better with a symmetrical filter... (just one of infinite possible examples). Then another addition to the mix, peoples' listening preferences are going to differ as well: for some, accurate tonality may rule, for others, perhaps transient impact may rule...

[sandyk]

The problem with such analysis is that, when considering 16/44.1, any approach to filtering is going to be compromised. That is, there is no "perfect" filter, there are only trade offs. So then the filter developer has to decide which parameters are more important for sound quality.

So, now we see that we are in the realm of purely subjective analysis: is reduction of pre-ringing more important? is extended frequency response more important? is reduction of overall ringing more important? is complete suppression of alias products more important? Who really knows? What we need is more study into psychoacoustics, where, maybe, some of these questions can be answered. But it will not be easy!

Then add the fact that everyone's "reference" system is different, such that some systems may sound better with a minimum phase filter with no pre-ringing, and other systems may sound better with a symmetrical filter... (just one of infinite possible examples). Then another addition to the mix, peoples' listening preferences are going to differ as well: for some, accurate tonality may rule, for others, perhaps transient impact may rule...

 

A welcome injection of a more general perspective. There is far more to the overall picture, and we still have so much to learn in MANY areas.