There are couple of intermediate phase (asymmetrical) poly-sinc filters available.
I know, but they are differentā¦
Personally, I tend to distinguish three general categories of sound characteristics that are HQP relevant:
-
precision and transparency : some filters (like the sinc, poly-sinc and gauss series) have a clear, precise and transparent presentation with a very low level of noise when combined with an aggressive dither. (Sinc-L is the paradigmatic example of this category, but then, it can also sound a bit harsh on imperfect recordings).
-
Weight and density, which I would consider to be the opposite of ādigitalā sounding, āmetallicā or āthinā. One filter stands out in this category: IIR.
-
Amount of āsheenā or echo, generally caused by the filter length or linearity (and also by DSD upsampling). I tend to prefer short or minimal phase, but although they help with voice separation, they sometimes sound a bit dry.
From a musical (rather than an audiophile) perspective, I tend to come back to IIR most of the time, because it conveys more emotion, it is richer and denser. Sure, it is less impressive than other filters, but the pleasure of listening is there. It let me forget about this audiophile hobby of ours and listen to music.
From my humble point of view, I think HQP already has a lot of āhi-fiā options to choose from. What I would personally like to see is more āanalogā sounding filters, especially more variations of this very unique IIR filter.
More importantly, I would like to have the ability to customize the inputs and output frequencies of HQP.
For example, like I just stated, one my favourite PCM filter, at least for now, is IIR. It sounds amazingly good and organically dense at 88kHz (from redbook), and a lot worse at higher rates (it somewhat loses weight and density). But then the fourth of my library is HD PCM/DSD content, meaning that I have to downsample. For some technical reason, 96kHz is downsampled to 48kHz, and 192kHz is downsampled to 96kHz. I like this filter so much that I accept to downsample everything to 88/96 including DSF files. (In so doing, the Gauss1 dither gives a bit more clarity that the NS4 one, although the NS4 is a bit less harsh). I wish I could instruct HQP to upsample 96kHz content to 192kHz, and 192kHz content to 384kHz.
(For the record, I used both a ESS and AKM chip dacs, and this filter PCM 88kHz IIR/NS4 combo is one of my favourite on the A26, which is the best DSD upsampling dac Iāve heard to date. For DSD512, I also enjoy Sinc-M, Poly-sinc-ext2, and Polys-sinc-short-mp a lot).
@jussi_laako, do you think that option could be on the roadmap of future HQP development?
Thanks!
1 Like
By the way, since you use A26 in PCM mode, does it have configuration options to select modulator and DAC chip mode? The chipset has four modulator options (just numbered 1 - 4) and two DAC configurations (āmeasureā and āsound qualityā). Just curious if DACs make these configuration options available.
In addition, the chipset supports 32x PCM (1.4/1.5 MHz), but Gustard at least doesnāt support this, AFAIK.
Iāll put it on my āthink about thisā list.
1 Like
Unfortunately, the implementation of the 4499EX chip on the A26 does not seem to include the four modulator options you mention (that would be for DSD direct mode, right?), but it does have the two configurations, they call them : āmonitoringā and ālisteningā (the latter seems to add a layer of polishing). I suppose that a) the āmonitoringā option is the direct route (so I use that in conjunction with HQP upsampling); b) I believe it is only applied in PCM mode, not in āDSD directā mode, which would be perfectly logical.
I know, thatās a pity! I also regret that Gustard didnāt use the latest third generation XMOS Bridge. But for now, it is still the best (of a few) options for this extraordinary chipset. I also own the E70 Velvet and for the price, it is quite good (nearly as good as the A26, quite frankly) although it does not handle DSD upsampling as well as the A26 does, for those who only do DSD upsampling.
Wow! Thanks! 
These are in the AK4191 digital filter / modulator chip. Not related to DSD Direct mode which it has. DSD Direct mode bypasses all this and goes to the AK4499EX chips with minimal processing in AK4191 (shortest possible path from data to D/A conversion).
In addition, the AK4191 has six different possible configurations for the DSD Direct mode.
OK, good, this has nothing technical specified in the datasheets. It just changes how the data is shuffled to the conversion elements.
Same goes for the modulators, absolutely no technical information, just four different modulator options (for going from PCM to the SDM data 4499EX needs).
1 Like
Thanks for recommending the IIR filter, @KML , I have not tried it yet but Iāll give it a shot. It sounds like it might be too āsmoothā to satisfy me, but Iām curious about how different it sounds from the other ones.
Iām growing more and more fond of the poly-sinc-guass-xla (extra-long, apodising) filter. I used to switch between gauss-long and sinc-L but recently have used sinc-L much more as the gauss filter just sounded too āpoliteā to me in the long run and music wasnāt stimulating me enough. sinc-L can, however, get under my skin for the opposite reasons - being always in-your-face type of presentation and also not fully convincing in terms of spatial cues (sounds deep etc. but I always feel like Iām watching a spectacle as opposed to participating in it).
Well, the gauss-xla really seems to hit the right spot for me. Itās like a denser, more vivid and punchier version of the gauss-long. At the same time, it clearly maintains the advantages that poly-gauss filters seem to have over the sinc ones. Iām hoping that this is the one that will stop me switching back and forth 
(maybe except for very short listening sessions where I just want to be āpunched in the faceā by Sinc-L, hehe).
As an example of the differences I was referring to and how I perceive them - if you listen to the beginning of Daft Punkās āFreshā, sinc-L might edge it out in terms of absolute depth but I feel disconnected from the scene being painted by the ambient sounds, whereas gauss-xla sends shivers down my spine, tricking my brain into thinking that Iām actually there.
3 Likes
What an Amazing album from Daft Punk.
Fresh! 
IIR: take, for example, the sustained note of the beginning, itās real and meaty. It sounds like an electric guitar, rather than a synthesizer.
Thatās precisely what got me into IIR! (If you try it, donāt forget to set the frequency output to x2, 88.2kHz from redbook. NS4 is easier on the ear, whereas Gauss1 is edgier).
Sinc-L : The transparency is absolutely immersive. Those waves! I want to jump into the sea. But it is fatiguing : I wouldnāt be able to sustain to the whole album like that. Plus, itās not grounded, I loose emotional contact with the music, it becomes an abstraction. (I prefer the totally different Sinc-M/Mx, which initially made my purchase HQP).
Gauss-XLa : very good at about everything except that I sometimes miss the analogic sound signature of IIR. (With ext2, it has a been a favourite all-rounder).
I totally agree. It is also denser than the Sinc-L.
Lol. Why would you stop doing that? Where would the fun be? 
(Iāve saved so many settings to easily jump from a sound signature to another. I just hope that I could select a set from Roon!)
That being said, I think a good criteria is to choose the filter that will make one listen not only to the whole song, but also to the whole album. If one develops listening fatigue, or loses contact with the music (if it becomes a hi-fi abstraction), than the filter is obviously not the one to choose from, even though it may sound amazing at first. Sure I love details and transparency, but not at the expense of musicality.
1 Like
This. The emotional connection is key. For me, sound is just a part of the experience. Itās when I can feel the music tug on my insides, is where I land.
1 Like
And where did you land?
With a 4090/13900k HQP Embedded, Roon Core on another pc, and hearing more of ASMD7ECv2@1024 with less stutters and with various filters, I still prefer AMSDM7EC 512+fs/poly-sinc-gauss-hires-lp @1024 which I could do with just 13900k. Debating on returning the card now.
Ha, interesting! My previous (and perhaps very subjective) conclusions where that itās always best to upsample in the native format. And for that, one does not need a GPU. But now that I have a proper DSD direct DAC (and a powerfull GPU) I might end up preferring PCM to DSD512 upsamplingā¦ More tests to be done!
This has more to do with the DAC implementation than anything else. If you have a DAC that will take DSD and does not manipulate it, DSD is the way to go with PCM and DSD content . I am sure that @jussi_laako can explain it better than I could so I will leave it to himā¦
1 Like
Unless you have one of the rather rare R2R ladder DACs, PCM is not native for the DAC and needs to go through whole bunch of DSP before conversion to analog.
For any delta-sigma DAC (all current production audio DAC chips), thereās always conversion from PCM to SDM (DSD-like) data for the conversion stage. This requires upsampling to several MHz sampling rates. Side effect is that since those ~$10 chips have very limited DSP resources, so they cut many corners in this area. For example digital filters are typically limited to just 8x. And rest is then something like copying each sample N times to increase rate before the modulator.
While if you run the whole process in HQPlayer, you can run full proper digital filters up to 1024x and then much more complex modulators before sending out the final bits for D/A conversion.
So it all boils down whether you prefer the DSP algorithms inside the DAC or the ones HQPlayer can provide.
Same goes on ADC side for recording. All the current audio ADCs are SDM type and PCM is result of elaborate conversion from DSD-like data to PCM. That eventually needs to be converted back to what it got started with for playback. This is the original idea behind DSD - to skip these extra SDM ā PCM ā SDM conversions. DSD64 was just too low, partially because they had to fit it on what was essentially a DVD data disc - in both stereo and 5.1 multichannel. Now DSD256 can finally deliver full performance of the original idea.
6 Likes
Thank you for the explanation, @jussi_laako. Iām certainly not the only one to be grateful for the time you take to answer everyoneās questions.
I know DSD upsampling is a somewhat sensitive debate among audiophiles and that itās been discussed everywhere. I wasnāt my intention to open that pandora box once more, but since someone else didā¦
From my limited experience, for every ESS or AKM based dac I heard in the past (besides the newest A26), the conversion from PCM to DSD always took something away from the sound (micro details, transparency, mostly). I came to the following conclusion:
-
either the DSD paths in those dacs werenāt the shortest route, as advertised;
-
or the conversion from PCM to DSD itself altered the sound (especially when changing family rates, although I donāt do that unless the dac only supports 44k DSD).
-
Maybe PCM and DSD offer inherently different sound signatures. Surely, there is something I fail to understand, but at least, I know that I hear a difference.
I found this diagram from AKM (I donāt think itās the latest chip though). I donāt know how to interpret it correctly: I can see two separate pathways, one for PCM and one for DSD. Is PCM being internally converted to DSD in the Modulator section?
I understand that the A26 cannot bypass the internal DSP of the dac in PCM mode. I use mostly the 5th filter from AKM named āsuper slowā which is supposedly a NOS-like filter.
From what I can hear, DSD on the A26 is not being altered in any way, and I no longer feel like DSD upsampling deteriorates the sound.
It is the first time I hear a dac in which PCM and DSD sound so alike. But I can still hear that upsampling from PCM to DSD polishes the sound a little bit (when comparing the same filter in both modes). PCM upsampling still sounds fractionally more detailed and transparent than DSD upsampling (Iām not sure that I prefer it, I just hear a very slight difference).
So, thank you for pointing out that the difference is most certainly attributable to the interaction between HQPās DSP and the dacās DSP. I am not sure what to think about the second or third possibility that I mentioned previously.
Edit: I did a quick comparison (some orchestral works by Prokofiev) between EXT2/LNS15 768kHz PCM and DSD512 EXT2/AMSDM7EC 512fs (on the A26), and I DO prefer the DSD version! Wow! I think I have much DSD listening to do!
2 Likes
To me, sending PCM to those DACs sounds just coarse and edgy. Just like looking at something made of Lego bricks. Missing certain final finesse.
Yes, the delta-sigma modulator there.
At the bottom you can see the āVolume Bypass DSDD bit ā1āā which is the DSD Direct (DSDD) mode. All recent AKM chips have this feature. However their conversion sections have changed a lot in couple of recent generations. So donāt expect all AKM chips sound the same. A26 has setting to enable this. And in addition there are two possible filters/arrangements for the data going to conversion section. This changes D/A conversion bandwidth.
Point of HQPlayer is to preferably replace all the DSP blocks there, which are āDATT Soft Muteā, āDe-empahsis & Interpolatorā and āDelta-Sigma Modulatorā. Which as you can see is possible.
Yes, it is not possible with any current production audio D/A chip. In PCM mode at most you get to replace the digital filter part, but not the other half which is the delta-sigma modulator.
Or is it just more coarse and rough? Not as close to a smooth analog waveform?
I would recommend ASDM7ECv2 modulator at DSD256 or DSD512 with this DAC. It will take quite a bit of processing power especially at DSD512 though.
I would also recommend to try poly-sinc-gauss-long and poly-sinc-gauss-xla for comparison with ASDM7ECv2.
6 Likes
Crystal clear, thanks, @jussi_laako !
I just heard it the same way as you do. I guess itās been a while since I gave DSD a proper chance. That might also explain why I preferred filters like IIR in PCM mode.
Thanks, will do!
2 Likes
With an intel i9-9900KF overclocked at 4,97 GHz and NVIDIA GeForce RTX 3060, I canāt get Gauss-long to play without stuttering (every 4-5 sec.) on DSD512 / ASDM7ECv2 setting. (It works fine with AMSDM 512+fs). I tried to change the buffer without success. I also tried grayed/checked combinations for both Multicore DSP and CUDA offload. Maybe there is a something else I overlooked that could work? If not, what would be the minimum CPU/GPU requirement for that?
Thanks!
My 9900K needs no GPU help and overclocking to output fixed DSD512 x 44.1kHz rate
1x = gauss-long
Nx = gauss-hires-lp
Modulator = 7ECv2
Using Embedded
Try without CUDA
And what is your CPU cooling like ?
Iām using the big boy Noctua
1 Like
Same!
44.1kHz does work at 512 / ASDM7ECv2. (with and without cuda). Thanks!
I wonder if you loose something when upsampling 96 or 192 pcm material to 44kHz DSD instead of 48k DSD? Is there any advantage to use 48k?