Experiments with digital filters and CCIF IMD

[JohnYang1997]

Topping dx3pro provides 6 different digital filters. Most people say they can't hear a difference. Sure the mechanics of the function of human auditory system don't really allow us to hear obvious difference of these filters. We hare most sensitive to frequency response. But we can't hear over 20khz so the difference there are not important to us. Then harmonic distortion is very common measurements for us to do. But real music are consistent of all types of harmonics and most likely our headphones have 100 times more harmonic distortion. Now that lead to the third possibility. IMD. Intermodulation distortions. There are already different ways of measuring numbers of IMD. But from my experience, they are far from accurate and won't show the whole spectrum (pun intended). So what I decided to do is to look at the frequency spectrum of different imd exciters.
(before I did the measurements, I did an extensive listening test, will talk about subjective experience later)
After I did different types of measurements from multitone, ccif, smpte, to din, aes dfd. CCIF and DFD show most difference between filters. And CCIF is just a little bit more obvious, and these two are essentially the same tests. So I'll just use CCIF for demonstration.

 

[JohnYang1997]

Firstly, all filters. As shown, there are definitely differences between them. I will show each of them and comment on them.

 

[JohnYang1997]

GROUP 1: worst
Before reading the decibel marking directly, the peak is at +65db So read accordingly.

Spoiler: filter 1 sharp roll-off

This one has many small interval modulation noises which make it one of the worst filters here.

Spoiler: filter 5 super slow roll-off

Filter 5 is super slow roll-off which similar to the NOS in square wave (almost perfect square wave response) but worst IMD here. The modulation distortion is just too much around 10k.

 

[JohnYang1997]

Group 2: good

Spoiler: filter 3 short delay sharp roll-off (default)

The default filter not too great not too bad. Looks like a minimum phase filter.

Spoiler: filter 6 low dispersion (or low dispertion according to a popular picture of AKM digital filters)

I don't quite understand what low dispersion means here but it seems very similar to filter3. It has a little bit of pre ring than minimum phase.

 

[JohnYang1997]

Group 3: better

Spoiler: filter 2 slow roll-off

This is one of the best filters here. Having a little bit more delay than filter 4 and as equal amount of pre ring as the post ring which are both very little. And the spectrum is also much cleaner.

Spoiler: filter 4 short delay slow roll-off

This is the best filter here. Short delay and does not has pre ring. The spectrum is just a little cleaner than filter 2.

 

[JohnYang1997]

Subjective evaluation.
Remember tbe difference I am hearing is not very consistent and comes in and out. Sometimes it's somewhat obvious sometimes there is no difference. This is the smallest difference that I have tried (kinda similar to hires vs cd quality , comes in and out.)
I listened before and after the measurements.
Before the measurements:
With string instruments, filter 5 is just a bit brighter at the top end. This is the most different one than others. I kinda prefer it in the beginning.
And then with some songs and at some point, filter 4 just sounds a little dull or dark. It kinda feels like what I heard from pcm1794 long time ago, which kinda pulls every thing down a little bit. Less floating imaging. And as I know pcm1794 has very low imd so I was leaning towards filter 4. (here I think filter 5 probably has too much imd) Then I tested some more different songs. I tested two songs from monster cat. The drum transient just feels a bit sharper on filter 2, filter 3 and 4 are kinda close. So I also lean to towards filter 2 a bit. But 3 is kinda not far off to both. Also probably because 2 3 4 are close together and 5 is already very different. I compared 2 3 4 the most. And liked 2 and 4 more. I did not have really much memory with filter 6 and 1. But I do hear difference sometimes after switching from 1 to 2. And 1 is kinda like 3 but possibly a little more dirty sounding. Idk why but I did not prefer filter 1. I could not decide between 2 or 4 to be my favourite. I could not really tell apart 1 3 6 either.
Then I did the measurements. Found out that 4 being the best and 2 is second. 1 3 6 are actually close in terms of peak amplitude but 1 has much more small noises in the intervals.
Then I went back and listened more. Now it kinda makes sense. 2 and 4 are the best ones. And low IMD sounds less lively. The imaging will move down a bit. Filter 3 is kinda dry and sounds like ol dac I remembered (same filter). It is a little bit more solid sounding than others. Like ERNN ENN. SOLID. Filter 1 is little bit more dirty and congested sounding not as natural. Still no idea about 6 because it is right after 5 and before 1. Hard to tell difference.

Now it got me thinking, probably IMD is more pleasing sounding sometimes and in some ways? Possibly the why many audiophiles like tubes. It add liveliness to the music. Is there any correlation to the reproduction and live music? Does room add IMD to the sound? I need more experiments. For now I am happy with filter 4 for my dx3pro.

 

[andreasmaaan]

Interesting work.

I'd suggest also taking frequency response measurements, as I'm not convinced changes in IMD of the degree we're looking at here can account for the differences you hear.

I'd also suggest you do some ABX comparisons between pairs of filters - perhaps start with the filters that you believe sound most different - to ensure that you actually can hear differences between them when psychological factors are removed.

 

[JohnYang1997]

Interesting work.

I'd suggest also taking frequency response measurements, as I'm not convinced changes in IMD of the degree we're looking at here can account for the differences you hear.

I'd also suggest you do some ABX comparisons between pairs of filters - perhaps start with the filters that you believe sound most different - to ensure that you actually can hear differences between them when psychological factors are removed.

It's really easy to do abx here. I just don't look at the panel and keep switching. This way I can only hear filter 5. It's somewhat obvious with right song (string instruments). I can't do this with others. It's too small of difference to do abx.

 

[daftcombo]

What is the look of the impulsion & FR for filter 5?
How do the strings sound?

 

[andreasmaaan]

It's really easy to do abx here. I just don't look at the panel and keep switching. This way I can only hear filter 5. It's somewhat obvious with right song (string instruments). I can't do this with others. It's too small of difference to do abx.

That would seem to be pretty strong evidence that there is no audible difference between all filters other than Filter 5. Whether this is due to increased IMD, frequency response, or phase response is unclear, but based on what we know about psychoacoustics, I'd hazard a guess that the most likely factor is frequency response.

 

[andreasmaaan]

Something else you could try would be to:

• record the output of Filter 5 and say Filter 6 (which appears to be a linear phase sharp roll-off filter based on your description, i.e. closest to ideal) through an ADC with sufficiently high sample rate (e.g. 88.2kHz or higher) and sufficiently low distortion
• use a minimum-phase filter to roll off the frequency response of Filter 6 so that it matches that of Filter 5
• ABX both files at the maximum possible sample rate you can using Filter 6

If the two files are indistinguishable in these circumstances, the change you are able to detect reliably via ABX when switching between the filters is almost certainly down to frequency response.

 

[Veri]

Interesting measurements though. Good job. Weird that Fast roll-off seems badly implemented, while short delay fast and short delay slow are much better (on paper).

At the very least, AKM's default seems like a good default choice I HIGHLY dislike AKM's naming for these filters though. Calling Super Slow / NOS 'natural tone' is just stupid. JDS Labs even defaults to the super slow / natural filter I think, which is a shame.

Short delay fast or slow seems like the way to go! Wouldn't have known without your verification @JohnYang1997 cheers

 

[JohnYang1997]

Something else you could try would be to:

• record the output of Filter 5 and say Filter 6 (which appears to be a linear phase sharp roll-off filter based on your description, i.e. closest to ideal) through an ADC with sufficiently high sample rate (e.g. 88.2kHz or higher) and sufficiently low distortion
• use a minimum-phase filter to roll off the frequency response of Filter 6 so that it matches that of Filter 5
• ABX both files at the maximum possible sample rate you can using Filter 6

If the two files are indistinguishable in these circumstances, the change you are able to detect reliably via ABX when switching between the filters is almost certainly down to frequency response.

I have rested till just now. And I could immediately hear the difference between filter 2 and 4. 2 sounds more open up (more highs and thinner) and 4 has more mids(fatter). It was very obvious for like 10 seconds (3 switches). And after 30 seconds, the difference fades away. I think there is really something to it. And the results just now is kind of consistent as before, 2 has more transient, 4 sounds a bit pulled down and less highs. I did have a chance to compare to others. But there is clear difference. Also I think auditoy system can learn and alter the processing(in the brain) to discern the different filters.

Edit: when I listen, I actually use 96khz sampling rate. This case there is no difference in frequency response (measured). I chose to use 48khz for measurements is because I want to make the difference larger (48khz showed higher difference because 19khz+20khz is closer to the 1/Nyquist frequency). So in fact when I am listening the difference is only the filters themselves and should be lower than measurements shown.

 

[JohnYang1997]

Interesting measurements though. Good job. Weird that Fast roll-off seems badly implemented, while short delay fast and short delay slow are much better (on paper).

At the very least, AKM's default seems like a good default choice I HIGHLY dislike AKM's naming for these filters though. Calling Super Slow / NOS 'natural tone' is just stupid. JDS Labs even defaults to the super slow / natural filter I think, which is a shame.

Short delay fast or slow seems like the way to go! Wouldn't have known without your verification @JohnYang1997 cheers

I believe that el amp uses sharp roll-off just like the dx3pro which is minimum phase and no prering. Super slow/nos looks like perfect square wave.

 

[daftcombo]

These filters?

 

[SIY]

I have rested till just now. And I could immediately hear the difference between filter 2 and 4. 2 sounds more open up (more highs and thinner) and 4 has more mids(fatter). It was very obvious for like 10 seconds (3 switches). And after 30 seconds, the difference fades away. I think there is really something to it. And the results just now is kind of consistent as before, 2 has more transient, 4 sounds a bit pulled down and less highs. I did have a chance to compare to others. But there is clear difference. Also I think auditoy system can learn and alter the processing(in the brain) to discern the different filters.

This is screaming for a real ABX test, no excuses or half-measures.

 

[JohnYang1997]

These filters?

Yes!

 

[mansr]

The digital filters you are dealing with cannot cause IMD. It is mathematically impossible. The variations you see are probably the result of different amounts of imaging which in turn produce different intermodulation products in the analogue parts of the DAC. To see this effect more clearly, play a single 23 kHz tone (with 48 kHz sample rate) and look for 2 kHz intermodulation products.

 

[Veri]

I believe that el amp uses slow roll-off just like the dx3pro which is minimum phase and no prering. Super slow/nos looks like perfect square wave.

It says here they default to 'natural tone' == Super Slow. Unless they misspoke and meant otherwise.
https://blog.jdslabs.com/2017/03/ol-dac-the-peoples-filter/

 

[JohnYang1997]

This is screaming for a real ABX test, no excuses or half-measures.

I think I'll post up 192khz recording of dac running 96khz, so you guys can hear for yourselves. I'll use gravedigger from architects and some other songs.