Audibility of digital reconstruction filters

[Jud]

Thanks for raising this topic, Dennis.

 

I think this is an appropriate place to raise a continuing question on filter behavior from another thread. The sequence is (1) my question; (2) Miska's response; (3) Fokus's response.

 

Something that interests me is what appeared to be a mild disagreement between Miska and Fokus in comments in this thread and the 192/48kHz one (though I could easily be misinterpreting).

 

Miska I think is looking at things from a time domain point of view and saying filter ringing should be audible as smearing of transients. Fokus I think is looking at things from a frequency domain point of view and saying for people like me who can hear to perhaps 15-16kHz, ringing should not be audible.

 

Am I accurately restating what each of you has said? Any further thoughts or comments?

 

There are two aspects, one is audibility of the ringing as such, and another thing is it's implications as transient smear. The smear length of typical 1 ms is equivalent of one cycle of 1 kHz sine wave, while the ringing frequency itself is around the Nyquist frequency. It rings for many cycles - the overall smear time is length of the entire ringing period.

 

Like in a video, you may not notice content of a single frame, but if there's "motion blur" (content of previous frames leaking to next one) over 25 consequent frames you will certainly notice it. If you think that a frame coming out now still would have 1% of content from one second ago mixed in. If you have a stationary image, you don't notice anything, but if you have fast movement, it becomes apparent.

 

I am always speaking of and thinking in both domains, as they are the same.

 

Take a linear phase filter with pre-ringing, once more.

 

If the filter cut-off is in the audible range then the pre-ringing is audible as a separate artefact.

 

If you move the filter cut-off frequency upwards then this audible artefact disappears.

 

If someone can provide me with sample music(*) files of two >20kHz filters that are identical, except in their amount of pre-ringing and that can demonstrate an audible difference to a majority of non-naive listeners then I would welcome that. I have been looking and trying for ten years, with listener panels, but so far ... nothing. So I lost all interest, I cook my filters pragmatically (just as Miska, sacrificing some above 18kHz to get to a shorter filter kernel, and always ensuring sufficient alias suppression - yes, I only care about the ADC side of things), and I carry on, concentrating on things that really matter.

 

(* Or even just a killer test sound.)

 

Miska and Fokus agree the ringing itself will not be audible for sufficiently high frequency filter cut-offs. But Miska goes on to say there will be an audible effect of the ringing, consisting of smearing of transients over a relatively long time compared to the ringing frequency (~ the period of a 1kHz sine wave). Fokus does not specifically comment about transient behavior, but does say frequency domain and time domain are the same.

 

So is Miska's remark about audible transient smearing something that others in the field can agree with? Do you instead think it is perhaps measurable but not audible? Or is it something you do not even see in the measurements?

 

Another interesting thing to me is Fokus's remark about sacrificing some response above 18kHz. Is that to get so-called "apodizing" behavior? I note in this connection that if I recall and understand his remarks correctly, Barry Diament, the producer, does not like what he hears from apodizing filters. (Barry likes the Metric Halo ULN-8. Anyone know what type(s) of filter(s) it uses?)

[Jud]

Okay, taking your theory as true, what accounts for the fact that different filters produce a different sound? After all, the only difference between filters is post, pre, or no ringing, at least as far as your experiment and derived theory go.

 

-Paul

 

?? I think the response was simply concentrating on ringing behavior, not saying that was the only difference among filters.

[Jud]

I want to concentrate on the audibility of different filters built by experienced people. That's simply because it's quite possible whatever combination of parameters I didn't like the sound of in iZotope SRC (or what I *did* like the sound of) is something no pro would ever do. I would expect less obvious variation in the sound of filters designed by pros.

 

That being said, I was floored by the degree of audible variation among the filter options provided in Miska's HQPlayer software. I would urge anyone curious about the topic of this thread to try it.

[Jud]

If someone did an impulse measurement on these things, that would be really interesting.

 

I have an idea for experiments along these lines, but I would like to run it by some people privately first.

[Jud]

Dennis, the differences between filters in HQPlayer, in non-blinded listening, seemed far more audibly obvious to me than I had assumed they would be. (I can think of blinded tests to verify this, but don't know if Miska would have the time or inclination to play along.) That's why I urged people to try it, because I think it might help reset skeptical expectations to something more like neutrality - "Hmm, think I hear something, do I really, and if so what could cause it?"

[Jud]

We want to be careful not to let incomplete models limit our picture. Inharmonic attack transients may be an important part of instrumental and vocal sound. It's also a neurobiological fact that different neurons process transient versus tone information. So though we have a mathematical model that can translate perfectly between tone and time, that is not the way all sound behaves, nor the entire description of the way we hear.

[Jud]

Testing with pure sines is completely useless, because this is not about hearing tones, this is about hearing waveform shape of rising transient. This is because transients naturally alert humans about incoming danger and that's why human hearing has developed to be especially good at detecting transients, rather than steady tones. Brain is trying to automatically filter out steady tones (background noise) to detect any changes/transients.

 

So you rather need to compare band limited and band-unlimited snaps, pops and crackle where the spectrum spreads across from low frequencies to high frequencies.

 

Miska: This makes logical sense (points in the same direction as some of the comments about pre-ringing due to filters), but it in essence asserts that we are separately sensitive to frequency and rise time through two different mechanisms. How would you conduct a test of the ability to distinguish between two levels of transient attack while holding frequency constant?

 

Already been done in the academic literature.

[Jud]

To the extent that "transient attack" and frequency response are treated separately, wouldn't the things you raise here similarly apply in terms of amplifier design/capabilities and speaker design/capabilities?

 

Also, if our brains process these two separately/differently, wouldn't it imply that some of us are more sensitive to one than the other and that might explain both our equipment preferences and the differences expressed about the audibility/inaudibility of certain changes?

 

The entire chain would be implicated, which of course raises the inquiry as to how capable the rest of the chain is of reproducing these attacks, and the degree of variation that might be found among components.

 

Re sensitivity, all the academic papers I've seen certainly show considerable variation among individuals.

[Jud]

So you are not acquainted with the school of thought that looks at the temporal spread of the highest ERB in the cochlea, and derives from this a requirement for 250 us, or a transition band of 4kHz?

 

When translated to midrange frequenties the same reasoning yields filters with demonstrably inaudible preringing. Interesting, not?

 

Thanks for the very interesting information. It makes a lot of sense to me that the frequency sensitivity of the ear is higher for continuous sine waves than for transients. To put it differently, to be sensitive to 50µs transients, the ear would have to be sensitive to sine waves - or wavelets - of much higher frequencies (since this transient is a linear combination of such higher-frequency sine waves or wavelets).

 

It may make sense, but it doesn't comport with the (small amount of) academic literature I've read. Fokus, are there some references you can pass along?