Thoughts on software based filters/ modulators/ upsamplers - have you tried them with your hardware and what pros and cons ?
Who uses Software filters/modulators/upsampling?
- Thread starter Audiophile Neuroscience
- Start date
Just to clarify the question, do you mean using software based filters/modulators/upsamplers to override your DACs’ own filters/modulators/upsamplers?
Because all DACs have their internal hardware filters/modulators/upsamplers. Even for an NOS (non-oversampling) R2R ladder DAC, if the DAC natively runs at 352kHz or 394KHz (8fs) and you’re playing 16-bit 44.1KHz, you’re still running a sample and hold filter where each 16-bit sample plays 8 times until you move onto the next sample.
Or similarly, if you’re playing a 2.82MHz (1fs) DSD and your DSD DAC natively has 16 DSD modulators, your DAC is not playing the same DSD signal natively to each modulator at the same time. Native DSD playback is usually running a shift register so each modulator is playing the DSD signal shifted by 1 sample.
Since most DACs are hardware limited in computational power, the pros of running software filters/modulators/upsamplers is that you can get better noise shaping and more precise upsampling for the modulators of your DAC in theory so you should get more accurate transients and lower noise leading to better timbral accuracy and instrumental separation. The cons is that because you’re running more traditional PC hardware, it is very easy to introduce significantly more noise into the DAC which leads to harsher sound, more lower level noise, worse timbral accuracy and worse instrumental separation. Another con is that you have infinite number of filters/modulators/upsamplers to choose from so you can easily get lost in things or you might prefer one filter that is less optimal for one recording and then ended up choosing a less optimal filter for your DAC. And if you don’t understand how your DAC works, you could just be doing busy work as you can feed your DAC a signal that your DAC would just turn around and run it through its internal hardware, negating most of the benefits of the software filters/modulators/upsamplers.
Because all DACs have their internal hardware filters/modulators/upsamplers. Even for an NOS (non-oversampling) R2R ladder DAC, if the DAC natively runs at 352kHz or 394KHz (8fs) and you’re playing 16-bit 44.1KHz, you’re still running a sample and hold filter where each 16-bit sample plays 8 times until you move onto the next sample.
Or similarly, if you’re playing a 2.82MHz (1fs) DSD and your DSD DAC natively has 16 DSD modulators, your DAC is not playing the same DSD signal natively to each modulator at the same time. Native DSD playback is usually running a shift register so each modulator is playing the DSD signal shifted by 1 sample.
Since most DACs are hardware limited in computational power, the pros of running software filters/modulators/upsamplers is that you can get better noise shaping and more precise upsampling for the modulators of your DAC in theory so you should get more accurate transients and lower noise leading to better timbral accuracy and instrumental separation. The cons is that because you’re running more traditional PC hardware, it is very easy to introduce significantly more noise into the DAC which leads to harsher sound, more lower level noise, worse timbral accuracy and worse instrumental separation. Another con is that you have infinite number of filters/modulators/upsamplers to choose from so you can easily get lost in things or you might prefer one filter that is less optimal for one recording and then ended up choosing a less optimal filter for your DAC. And if you don’t understand how your DAC works, you could just be doing busy work as you can feed your DAC a signal that your DAC would just turn around and run it through its internal hardware, negating most of the benefits of the software filters/modulators/upsamplers.
Last edited:
Many thanks for detailed reply
I have the Gryphon Kalliope DAC with "selectable32-bit/210 kHz asynchronous sample rate conversion and a dedicated ESS SABREES9018 32-bit D/A converter per channel, incorporating eight individual D/Aconverters in Dual Differential coupling"
You hit on an interesting point about "overriding" the DACS filters/modulators/upsamplers. I have only a rudimentary knowledge of such things but can understand that a digital signal can be upsampled prior to sending to DAC, and presumably delta sigma modulation from a multibit digital number to high speed 0 and 1s can also occur prior to the DAC and presumably filters like slow and fast, linear phase etc,apply to that SDM. Presumably the analog reconstruction filter can only occur in the actual DAC. If correct, yes I get that using a really powerful computer could be advantageous at taking these tasks out of the hands of the DAC, and you can add many more choices for better or worse.
I also understand that using PCs may introduce noise although they can be built for purpose with low noise PSUs and components
Still, what strikes me as odd is how does this new improved signal play with your DAC. Do you turn off oversampling for example etc
The next question would be, in the case of expensive DACs, one would expect that the computational power has been matched to the required functions
Then, as you say there may be a bewildering number of software settings and some may work better or worse depending on different circumstances or clash with your DAC
Not the least, there is also the spectre of PC and software glitches, learning curves, and many hours spent on forums to fix potential issues which can divert you from enjoying the music.
So, my original question is in an attempt to find out how people here view software based media players stacked with filters/modulators/upsamplers. I am also curious what the experience is in relation to higher end DACS which one would hope would perform well in the first place.
Last edited:
Across multiple systems and DACs I’ve experimented with HQ Player’s upsampling (inline) and also PGGB (offline).
When the settings are “just right” there has been, in many of my setups, a noticeable improvement in “air”, “space” and lack of harshness. Settings have varied by DAC and even for a given DAC by the bit rate of the source material. The author of PGGP, which I apply offline, has offered great advice on settings to avoid the endless experimentation.
There’s a long PGGB thread over on Audiophile Style.
IMO it’ worth the small price to play with it, if only to better understand the sonics of alternative algorithms.
When the settings are “just right” there has been, in many of my setups, a noticeable improvement in “air”, “space” and lack of harshness. Settings have varied by DAC and even for a given DAC by the bit rate of the source material. The author of PGGP, which I apply offline, has offered great advice on settings to avoid the endless experimentation.
There’s a long PGGB thread over on Audiophile Style.
IMO it’ worth the small price to play with it, if only to better understand the sonics of alternative algorithms.
I have a Gryphon Diablo 300 with internal DAC, which I think is similar to the Kalliope DAC.
I also have a Holo May L2 R2R DAC.
I use HQ Player running on a Mac Mini on my network, the data connection to the streamer is fibre optic. The streamer is Innuos Pulsar, which has HQ Player Endpoint embedded onboard.
The Gryphon Diablo 300 DAC is improved by sending data upsampled to 32/384 PCM. There is a distinct improvement in resolution.
Using Holo May in non-oversampling mode with HQ Player upsampling to DSD256, there is more detail and air, but in comparison the Gryphon Diablo 300 DAC has a little more warmth and is pleasing in its own way.
Neither have any audible harshness or noise whatsoever.
The thing about the Holo May in a blind test with a Lumin was the increased air and resolution because the noise floor is about 20dB lower. Inaudible noise floors matter!
I hate the expression, by the Holo May is a touch more analogue, even though the 300 DAC is still very good. So I kept the Holo May.
So I can comment more on some questions on software processing later.
But for @Audiophile Neuroscience the bigger picture question is: Is it worthwhile to build and tweak a low-power low-noise computer to play with these filters into the Gryphon DAC which is a 2014 product, or is it better to just get a new, maybe even cheaper 2024 DAC which might or is likely to outperform the Gryphon. I personally like the Chord DAC sound and thinks the Qutest is great even though the DAC is finicky to setup for optimal sound. You heard from others that they like their Holo May or MSB sound.
Sure you may be able to eventually use software upsampling/filter/noise-shaping to improve any DAC, but I think it’s only when you’ve finished upgrading your DAC before people should explore software solutions for further upgrades.
But for @Audiophile Neuroscience the bigger picture question is: Is it worthwhile to build and tweak a low-power low-noise computer to play with these filters into the Gryphon DAC which is a 2014 product, or is it better to just get a new, maybe even cheaper 2024 DAC which might or is likely to outperform the Gryphon. I personally like the Chord DAC sound and thinks the Qutest is great even though the DAC is finicky to setup for optimal sound. You heard from others that they like their Holo May or MSB sound.
Sure you may be able to eventually use software upsampling/filter/noise-shaping to improve any DAC, but I think it’s only when you’ve finished upgrading your DAC before people should explore software solutions for further upgrades.
I believe Sabre DACs are 32 or 64 element SDM. I don’t know what frequency they run at but let’s say it’s 2fs DSD so 5.6MHz.
Traditional chip DACs generally has two components. The first is that there is an upsampling filter so your 44.1kHz CD music is often upsampled to say 352kHz 24-bit. The more computational power you put into the upsampling filter, presumably the more accurately you can recreate the original analog waveform when it was recorded, leading to better instrumental separation, low level details, fuller bass and better timbral accuracy. For me the thing that jumps out is when an orchestra is playing, good upsamplers can help you easily pick out individual instruments whereas most DACs would help you hear the main melody but you have to strain to hear the harmony and orchestration, particularly if the instruments are playing softly.
Moreover as pointed out earlier, your DAC actually has 64 elements of 5.6MHz 1-bit output so the DAC needs to upsample and convert and map 352kHz 16-bit to those 64 elements of 5.6MHz 1-bit. This generally involves some form of noise shaping. The more computational power you throw at noise shaping the more accurate the small signal accuracy and the lower the noise floor in the audible range. This can give deeper soundstage, more micro details, more dynamic contrast. A good noise shaper paired with good DAC design can also minimize noise floor modulation causing less harshness or brightness.
The reason why software products can do a better job with upsampling filters and noise shaping is because modern GPUs are very powerful and made with modern 5-8nm silicon fab. And they can do Fourier transforms to accelerate the computation. By comparison, most DAC chips are made on much old silicon fabs and don’t have that much computational power included and upsampling filters often require direct convolution. To efficiently do these calculations in hardware, you’ll either need to pre-program a DSP chip or an FPGA chip which can also get very costly. Hence even for most DAC products with FPGA chips in them, the hardware actually does very little computation. I think only Chord and MSB actually program their FPGA to do a significant amount of computation as evidenced by the audio delay. I believe there is at least one company out there that makes a high performance DSP chip that some high end audio manufacturers use for their DACs. But since so many things are trade secrets and proprietary, I’m not sure.
Specifically to Gryphon Kalliope, I suspect you can use an upsampler and feed it 352/24 PCM signals and get better sound as the signal would just go through the noise shaper to create the sound.
Some DACs when you feed it DSD would play natively (with shift register) so all 32/64 elements would play the same DSD signals you send it. Others would convert the DSD to 352/24 PCM first and then feed that to the noise shaper. I have no way to know what gryphon does. If it’s the former, then you can use software to send 2fs DSD to Gryphon. The problem with 2fs DSD is that it doesn’t allow for much noise shaping because 1-element 1-bit 5.6MHz is actually quite low resolution. This is why people who like to do software processing get DACs that can handle 8fs or 16fs DSD (22.4MHz or 44.8MHz)
Traditional chip DACs generally has two components. The first is that there is an upsampling filter so your 44.1kHz CD music is often upsampled to say 352kHz 24-bit. The more computational power you put into the upsampling filter, presumably the more accurately you can recreate the original analog waveform when it was recorded, leading to better instrumental separation, low level details, fuller bass and better timbral accuracy. For me the thing that jumps out is when an orchestra is playing, good upsamplers can help you easily pick out individual instruments whereas most DACs would help you hear the main melody but you have to strain to hear the harmony and orchestration, particularly if the instruments are playing softly.
Moreover as pointed out earlier, your DAC actually has 64 elements of 5.6MHz 1-bit output so the DAC needs to upsample and convert and map 352kHz 16-bit to those 64 elements of 5.6MHz 1-bit. This generally involves some form of noise shaping. The more computational power you throw at noise shaping the more accurate the small signal accuracy and the lower the noise floor in the audible range. This can give deeper soundstage, more micro details, more dynamic contrast. A good noise shaper paired with good DAC design can also minimize noise floor modulation causing less harshness or brightness.
The reason why software products can do a better job with upsampling filters and noise shaping is because modern GPUs are very powerful and made with modern 5-8nm silicon fab. And they can do Fourier transforms to accelerate the computation. By comparison, most DAC chips are made on much old silicon fabs and don’t have that much computational power included and upsampling filters often require direct convolution. To efficiently do these calculations in hardware, you’ll either need to pre-program a DSP chip or an FPGA chip which can also get very costly. Hence even for most DAC products with FPGA chips in them, the hardware actually does very little computation. I think only Chord and MSB actually program their FPGA to do a significant amount of computation as evidenced by the audio delay. I believe there is at least one company out there that makes a high performance DSP chip that some high end audio manufacturers use for their DACs. But since so many things are trade secrets and proprietary, I’m not sure.
Specifically to Gryphon Kalliope, I suspect you can use an upsampler and feed it 352/24 PCM signals and get better sound as the signal would just go through the noise shaper to create the sound.
Some DACs when you feed it DSD would play natively (with shift register) so all 32/64 elements would play the same DSD signals you send it. Others would convert the DSD to 352/24 PCM first and then feed that to the noise shaper. I have no way to know what gryphon does. If it’s the former, then you can use software to send 2fs DSD to Gryphon. The problem with 2fs DSD is that it doesn’t allow for much noise shaping because 1-element 1-bit 5.6MHz is actually quite low resolution. This is why people who like to do software processing get DACs that can handle 8fs or 16fs DSD (22.4MHz or 44.8MHz)
I converted to Hqplayer a couple of years ago. I convert all local files/Qobuz on a dedicated Linux based PC downstairs and this feeds DSD512 to my main system upstairs over the network. The endpoint (termed NAA in the Hqplayer ecosystem) is a small low power fanless PC that sits behind the audio rig. This feeds either a Lampi DAC or a Holo Spring L2:V3.
Last edited:
Fair point regarding Kalliope now 10 years old . IIRC there was supposed to be some kind of upgrade path for software filters and/or other aspects. Probably worth exploring.
That said, one gets the impression that software programs run on a computer can almost bypass the DAC all the way to the analogue filter stage? If so, it may makes less sense to buy a new DAC especially since I have a dedicated pc based hdplex server with linear power supplies and dedicated usb card etc (albeit 10 years old also).
I also assume that there is more to any DAC than the DAC chip, as in implemetation, topolgy, power supplies etc
Thanks
the kalliope has I have the Gryphon Kalliope DAC with "selectable32-bit/210 kHz asynchronous sample rate conversion and a dedicated ESS SABREES9018 32-bit D/A converter per channel, incorporating eight individual D/Aconverters in Dual Differential coupling"
so is that now 32 bit? The frequency I think is 384kHz. I am not sure what 64 elements means.
The Mojo Mistique X DAC referred to above uses 40 year old chips. I spent a couple of months listening to the 8-10 year old ESS9018 based Gryphon DAC feeding it 32/384 from HQP and it was very good. Could happily live with it.
You appear to be using your computer as the transport, so you have a usb connection directly to the DAC. I go via an Innuos streamer fed by fibre optic, then usb to the DAC.
The key parts in the Holo May DAC are the usb input and the analogue filter. The no-frills Cyan 2 DAC includes the required elements and costs a little over $1,000. Seems to have been optimised for desktop usb use with HQP. They included usb galvanic isolation and excluded onboard oversampling. It also runs cold, whereas the Holo May runs quite warm.
The only written review I found so far was on The Ear and he used Audirvana rather than the much better HQP upsampler. A very Slavic online review said very good in NOS mode, better with HQP upsampling, but stick to native format. He thought PCM to DSD reduced dynamics.
Anyway, I don’t think there is much dispute that HQP is the best software upsampler (if not the easiest to implement) and Holo is its best partner. Holo now have a bare bones DAC focused on usb sourced HQP, for less than some here spend on the USB cable. I fancy getting one just to see how good it is.
Incidentally, this is the networking and processing department, in a wall cavity behind a bathroom. It has a dedicated spur using Belden drained mains cable. I always use a continuous power supply for computers, CyberPower APC in this case.
The Ubiquiti network uses a controller (little white box) to create a VLAN for the hifi and a 24 port and 2xSFP managed switch with PoE powering 5 access points (not used for the hifi). The connection to the hifi is fibre out of that switch.
Top row:
The Mac Mini M1 runs HQ Player and Roon Server.
Network controller
The tall black box is a fibre modem.
Finally, a Buffalo SATA network drive containing my music files.
Below:
Managed switch
QNAP TS-473, was used for Roon Server, now on domestic and work storage duties.
APC
The Ubiquiti network uses a controller (little white box) to create a VLAN for the hifi and a 24 port and 2xSFP managed switch with PoE powering 5 access points (not used for the hifi). The connection to the hifi is fibre out of that switch.
Top row:
The Mac Mini M1 runs HQ Player and Roon Server.
Network controller
The tall black box is a fibre modem.
Finally, a Buffalo SATA network drive containing my music files.
Below:
Managed switch
QNAP TS-473, was used for Roon Server, now on domestic and work storage duties.
APC
The 32-bit/210kHz ASRC is just another way to filter the incoming signal. It is not a computational intensive filter. The primary reason to convert to a non-multiple of original sample rate is mostly to reduce jitter.
And yes, the DAC is capable of 32-bit. I should have said internal upsampling to 32-bit 352kHz or 32-bit 384kHz. Usually upsampling is an integer multiple (hence number x fs). To simplify (maybe over simplify), most DACs are just multiple 1-bit DSD/SDM running at higher frequencies combined together. Each of these 1-bit DSD/SDM is called an element.
Interesting. I thought PGGB is better for offline upsampling based on what I’ve read online. I have only tried multiple early versions of HQP and PGGB and I felt they were inferior to my Chord setup. But I did feel at the time PGGB was better than HQP. Some people tell me the newer versions with newer filters are better but I don’t like to be HQP/PGGB’s beta tester (and definitely not a paying beta tester) so I have not re-tried them for a very long time. So I don’t feel I have any real insights on the current versions of HQP or PGGB
My “cloffice” downstairs- Ubuntu/HQP main PC. I have since disabled the RGB LED:
My Fitlet2 behind the system upstairs - the usb has the NAA boot image, I have since loaded on a micro SD card for lower power requirements. Teddy Pardo PS is barely shown on the left:
HQP effectively has very good support through Roon Community.
HQP is a lot cheaper compared to PGGB Max (about a quarter of the price)
Innuos Sense includes an HQP endpoint, which is an slightly optimised approach.
I think Chord favour PCM, Holo favour DSD.
I can't cope with two complex software packages at once.
I don't really understand computers, which is why I use Apple, which also work for me in business. Not used Windows for over 20 years. My Mini M1 was sitting around unused after being replaced in my desktop system with an M2 Pro. So my HQP set-up cost be $250 for the software and I added the VLAN controller for $130.
Running Roon and HQP on a Mini M1 uses almost 40% of the processing power, with nothing else but OSX running. If you are going to use upsampling, that seems to me an awful lot of processing power to have onboard a DAC. If you use an FGPA chip to do high rate DSD like PS Audio, you end up with a high noise floor. If you do it the dCS way, you get a big debit from your bank account. I don't like a high noise floor or big bank debits. Taking the processing off-board just makes more sense to me. Plus the data can be sent over fibre rather than the processing polluting your DAC unless you spent a fortune.
I did try a not inexpensive multiple choice upsampling DAC and was disappointed. Even after a retro fit Teddy Pardo PS.
Last edited:
Having the high heat/ noise/ processing separated is key - one of the geniuses of the Hqplayer NAA. Mine is a headless server and I can manage it via SSH from another PC on the network. The PC doesn't necessarily have to be headless though. Windows , Ubuntu desktop, and Mac OS work as well in the same configuration -as per your example.
I don't know what NAA is. I just loaded the software, did the settings I was told to do and switch the streamer to HQ Player endpoint. It says my streamer is NAA, which means nothing to me.
The new version of OSX that came out a few weeks ago allows you to be permanently logged in to a headless Mac mini on the network and switch to it, without having to go via Screen Sharer. My configuration is extremely popular for those without IT skills.
Similar threads
 | Steve Williams Site Founder | Site Owner | Administrator |  | Ron Resnick Site Owner | Administrator |  | Julian (The Fixer) Website Build | Marketing Managersing |