Hi Peter

Whenever it comes, I'm happily anticipating a 384KHz Custom Filter.  And even if it is beta, I'll gladly try*any one*.   This is because the 192 is not right at 384, with too much buzzing, and with ArcP being too much on the clean / dry side most of the time.   

Much has happened here with my DAC; I've had to return it due to that same initialization / synch problem, and so the fixed DAC works right  but I've had to break it in a second time and so it has been a moving target.   

I was previously running with the slow roll-off filter and now that has changed as I'm now back to the brickwall on the DAC.   It sounds better right now.  And believe it or not the best sounding filter within xx is back to the new "AI" with minimum phase settings and TPDF dither.  This could be because it is a preprocess or because it is at least tuned to 384 and maybe I just can't hear the buzzing because it is too high in frequency.  Along with this, Miska's asymFIR filter at 384 is the best sounding while xxHighEnd is still the most impactful by far.

But I highly suspect that any custom 384 filter in XXHE will be, both  the best sounding and most impactful. 

Bring it ON!!!!

In contrast, this album The Very Best of the Manhattan Transfer plays by far the best with Arc Prediction upsampling at the highest sampling rate I have available, 352.8.
It plays hyper cleanly with arc prediction, very analog, and the vocals sound lush without the help of any added ringing/reconstruction. 

With Custom (176 @ 384) it sounds very good, but if you focus on the vocals there is clearly some buzzing going on from unneeded reconstruction.  So this recording has plenty of post processing on the vocals to get a certain sound.

With the newAI (set to 0, TPDF) this album plays with clearly audible distortion in the vocals (only).  There is too much ringing / reconstruction going on.

----------------->  Not tied specifically to this album:

I've noticed with these three upsampling methods that the vocals climb in the mix with added ringing (more added harmonics) and fall in the mix with less.  In one extreme we have newAI with the most ringing and where I perceive the vocals as louder.  In the other extreme we have arcPrediction with no ringing where I perceive the vocals a little lower in the mix.  In the middle we have Custom.

What's interesting to observe between these methods is that, because the ear is most sensitive in the vocal range, I tend to adjust the volume based on the vocal's loudness or perceived loudness.  The minimum step on my dac for volume is 1db.  So in practice, after I adjust the volume to compensate for the vocals, it is the band / instruments that vary in the mix.   The end result is that with Arc Prediction I get the most "band" in the mix, with newAI I get the least amount of "band," and with Custom I get somewhere in between.

This is all very subtle.   Even though subtle, it can make a big difference in perception and overall experience.  With more band from arc prediction can come the  perception that arc prediction brings the best sound stage and imaging (which appears to consistently be true).  At the other extreme we have the newAI.  Custom comes in very close to Arc Prediction for perceived sound stage, usually.

I found two good examples of where newAI filter using 0 (min phase) and TPDF as its variables sounds much better than the new Custom.  I post these two titles so that you may try them for yourself if you happen to have this music in your collection.

The first is Traffic Far from Home.  This is in the class of popular or rock music.  Here, using Custom, Steve Windwood's vocals sound as if only midrange is present.  Assuming your system might reproduce what I'm hearing, it sounds almost as if he is singing through one of those hand-held "megaphones" or institutional outdoor speakers used for public announcements.  For sure the feeling that something is missing.  Now switch to AI.  With AI it sounds like a professional recording, proper, with an entire range of high frequencies blended in with the midrange.  It sounds like a real person singing.

Next is Antal Dorati and the London Symphony Orchestra, as released by Mercury Living Presence Léo Delibes Sylvia & Cappélia, Disc 2.  This is classical music and there are all sorts of different stringed instruments.  The focus here is on the delicate sound of the strings.  With Custom, it sounds good;  but a little bit like a CD, which it is, and so has a little bit of that digital hardness.   Now switch to AI.  With AI it sounds wonderful, like real strings.  Forget CD sound, it sounds like HD material or SACD.  There's just a beauty to the strings that is not there with Custom.  Folks,  it sounds like a live concert, very smooth, very analog, and with a very nicely balanced high end.  Further,  with this recording, there's no buzzing / zooming or distortion as I'd normally expect with the extra ringing.   

Both were played at volume levels that sounded best for each:  -3 db for Custom,  and 0db for AI.  (And for Traffic I think it was -4.5db for Custom)

Because it takes so long to load AI at 8x, I tend to use Custom.  So making the above realizations can take some time.

A simple hypothesis would be interaction, ..with the ringing already present in the recording.  

And so the most natural sounding interaction is the one that goes unnoticed.

This would also explain why it is so difficult to pick a favorite filter amongst, for example, Miska's 15 or so filters in HQPlayer.  All sound different, and all eventually lead to a situation where a change to a different filter improves.

My experience with MP3 is that is has a reduced, greatly reduced, soundstage.

Perhaps Spotify is assuming that listeners are using ear buds and they apply a binaural filter to simulate the soundstage.  There is much magic that can be performed with this.  I just bought a new car and the faux surround sound is great, with instruments all around.

I still use it with ArcP.  Yes, I've heard a difference.  It doesn't come into play that often if you don't use it, but playing with it on is the safer way to go,  IMHO.

This happening now fairly consistently and therefore often enough to report.  Regarding the newAI vs. the 176 custom, I don't know how this is possible but it appears that the 176 custom filter at 352 can have more buzzing / zooming than the newAI filter using 0 (min phase) and TPDF as its variables.   This is not supposed to be the case.   Very strange.

As always,  ArcP none, or whatever is in the recording.

[BTW,  if it ever mattered, only now is my Sig remotely close to what it has been for months and most of this thread,  except that wasapi is new within the last few weeks.  The one thing I'm quite certain is inaccurate is the Q5 setting; it might be 5]

Hi Peter

Sorry for the delay.  Busy week.

There's a bandwidth or slope setting,  I don't have it in front of me right now.  As I recall it shows Fs/2.  It appears to be clickable with a drop down, but it isn't.  I guess this is grayed out.   I was hoping I could change the value to something like Fs/1.8.   

What I'm really trying to do is open the slope or control the slope, either the starting point, the ending point, the rate, or all three.  One interesting one would be starting from 16.3K and ending at say 22.05.   Another would be starting at 20K and ending at 24K or 30K (under the assumption that the original AA filter at the mastering stage did its job, as an option anyway).   

Why?   Because to my ear 48KHz Fs recordings sound so much better than 44KHz Fs recordings.   48 can sound almost as good as 96 or 88, just shy, or sometimes indestinguishable.

And okay, other ratios may be better than these.

I acknowledge that "opening" the filters does not always work.  For example, DSD sounds best with the lowest setting in my DAC, denoted by "<50KHz."   Meanwhile the chip default curves are 50, 60, and 70 KHz and none of these are acceptable.   One reason is that the quantization noise in DSD is full scale.   Whereas in PCM the quantization noise is over the last few least significant bits.   It seems that PCM has the advantage for this type of added noise.

I've been trying many different settings for the new AI, including all of the dither options ...including none.

At first I was trying to get an asymmetrical filter to work out.  I was trying values between 17% and 5% pre to post ratio.  And while an asymmetrical filter can sound good,  the best "new AI" settings I've found is for a full minimum phase filter with straight TPDF dither.  The settings for this are 0% and then selecting TPDF from the dropdown menu for dither.

I wish the divisor for the filter bandwidth were active,  I'd try opening it further if I could.

And yes for dither straight TPDF is sounding better than any of the noise shaped variants.

This is regarding the 176 Custom Filter, run usually at 352.


Correction: The quoted text above (from me), while true sometimes, is not true most of the time.  It depends on the compression level of the recording.

I've been running three weeks now at the recommended -6db and xx has turned out to not be my favorite player in this time.  The sound became drier, even with the custom filter which had solved the dryness.  AI sometimes came to the rescue but takes too long to load to use every day. 

So yesterday, I came across a dry sounding situation that was only dry with xx and not with other players.  Hmmm.  This was on the highRes version of the Indigo Girls All That We Let In,  This recording I know has peaks that never go above 90%, and I was upsampling from 96 to 384.   So I changed the attenuator from -6db all the way to 0db.  The dryness was gone!!!   

The same is holding true for most recordings that I've tried since then.

And so it appears that at -6db there is less "ringing" going on, and less reconstruction of the higher harmonics, than at 0db.   

My standard setting right now is -1.5 db and if I hear anything funny I drop to -3 db or -4.5 as needed.  The more compressed the recording, the lower I must go.

An attenuation of no less than -6db has been recommended with the custom filter.  In my case I'm talking about the 176 filter, used mostly at 352 and also at 176.

For sure audible clicks are heard at 0db.  For a while I ran at -3db of attenuation because I'm running deep class A on my ams and due to heat I have limited headroom.   The purpose if this post is to confirm that -3db is not enough.  When I run at -3db I don't hear clicks, but it is possible to hear things that don't sound right.  Further attenuating to -6db solved that problem. 

It's still hard for me to believe that the intentional ringing is reaching with peaks of  200% which corresponds to -6db,  but it appears to be the case that 141% which corresponds to -3db is for real.   

I've not tried -4.5db and may never try it unless I actually need the headroom / volume.

Q3-5=1 1 1 @ clk=15ms sounded best of all last night.  But tonight I preferred my old setting of Q3-5=0 0 3 @ clk=5ms, which I arrived at independently of when and how you and others did.  Plus, with 0 0 3 the unattended background updates quickly, whereas with 1 1 1 it can take 30 sec to 2 minutes.  1 1 1 sounds cleaner and 0 0 3 gives a more open and richer sound over here.

So I  can only imagine that 1 1 1 is somehow resulting in even CPU usage, and even HIGH CPU usage, correct?

Let me clarify what I understood because I think things are different at 352/384 than they are at 705/768 and I think that even hi res needs help.  I understood that you might make a custom filter that behaves like AP when the source is hi.res. and like Custom (how it is now) when sourced with 44/48. 

What I'm saying is that even with hiRes source material, at 352/384 my ear is much preferring Custom over AP.  Hard to argue with the ear.

As an aside, I now hear the buzzing / zooming / mosquito-ing of the typical upsamplers that you've talked about.  I also hear that AP never has it, nor do the interpolative filters in HQPlayer.  With Custom, I can *sometimes* hear a little hint of the buzzing.  I accept it as it comes with a bunch of other goodies.  But usually when I hear it with Custom, it is in the recording itself; and I presume this is from the EQ processors that audio engineers use for vocals.  No doubt they ring, some more than others.   And I can check with AP for those recordings and sure enough the ringing is there for those recordings.  It's not a clean ring when it is in the recording but the essence is surely there.

Anyway, I hear the trade off that you are making with these filters and can see that it is like walking a tight rope, picking the right balance between the two extremes.

I can say, it is amazing that Custom measures better than AP because AP sounds cleaner.  But sound is sound.

I think that this is true for 705K.  But don't forget us peons who can only go to 384K! 

Thanks,  and...


Also, with regard to this, when I try 1,1,1 the task manager does not update.  Normally I look at the graphic showing all 12 virtual cores to see what is happening and where.  But the entire task manager pane just stays as it was before I started playback.  After I it stop it eventually becomes live again and I can see the alt-x activity a little bit.  So this 1,1,1 is a powerful tool for controlling resources but I can't see what's going on.