When I started on my original PCM upsampling filters for the DS prototype I also wanted to minimize “preringing” But that isn’t mathematically consistent with preserving the waveform shape which I believe is more important and which sounds better to my ears.
(On a slightly different subject) DSP textbooks teach that the exact factoring of anti-imaging filters doesn’t matter and that you can make huge savings in the processing power and memory usage with a “proper” factoring. Rob Watts, I, and very probably others have learned that the textbook approach doesn’t sound the best. I am gratified that there are a few papers out there that discuss this from a DSP accuracy perspective (i.e. not audio related) but most practitioners don’t believe it.
Trusting your ears definitely can lead a different direction than common practice.
I’d like to add that I don’t believe anything magic or unexplainable is happening. Unlike mathematics, in physics, etc. you are always working with models of the real world and one of the first things engineers should learn is to pick the best models for their purposes. As a small example
Many objectivists are misled by using a simple RLC model of cables which glosses over many things that are (sometimes marginally) related to their use for audio. Galen understands that more complicated models are needed for designing good audio (and other) cables and that shows in the measured performance and the sound quality of his cables.
The reason I like the papers I found about factoring resampling filters is that they use a more complete model (and a deeper analysis) of the numeric accuracy of the math used in picking the coefficients and implementing these filters and that result better matches Rob’s, my (and other’s) experience implementing such filters. I really wish more engineers had to take a quality numerical analysis course these days. The one I took about 45 years ago was taught by a mathematics professor that really didn’t trust calculators (he was burned by the HP-35’s trig functions) he taught much more complete analyses than the text provided.
When someone doesn’t necessarily trust common ground…or even proved technology…combined with the ability and knowledge of lateral thinking…something bigger is going on.
It seems you met a few people like that on your way to do it similar
Absolutely, whatever works and sounds best. I’m just a little greedy to have my PCM to sound effortless as DSD as much as possible! I know in most traditional DSD systems there is very little filtering at all, and often none on the chip (which I think sounds better than the ESS design). However, I am also starting to really appreciate that much of DSD’s sound quality comes from the low level resolving ability, that even 24 bit PCM doesn’t have. According to Michal at Mytek Digital, 32 bit integer PCM is a lot closer to DSD (even though the dynamic range is unchanged), but nobody is using it yet.
Well, 24 bits has the dynamic range to go from a single molecule hitting your ear drum to standing in a jet exhaust. More than 24 bits at the DAC aren’t necessary (but when editing PCM more bits are definitely needed for almost any manipulation.)
With current technology we don’t really achieve 20 or 22 bits of linearity, so 24 bits is still a great, but hard goal.
In addition to the inherent linearity of DSD I do think that the higher sample rate of DSD has a lot to do with its sound quality and transparency.
Back on subject, with every release I do try to increase the transparency of both PCM and DSD. There is still some work on upsampling that I intend to do as well as other places in the FPGA.
Hope it isn’t too off topic and too technical, but I think the idea of 32 bit audio is just with sigma delta chips the conversion of the input/output bitstream (11.2mhz 6 bits, for instance) to/from PCM is less lossy and has better filtering DSP when you use 32 bits instead of 24, and for some reason sounds better.
Just being pedantic, but you said “32 bit integer PCM is a lot closer to DSD”. I took that as a PCM format, not an internal representation during processing. I use more than 70 bits for the filter proper and more than 60 bits for the PCM to DSD conversion, 32 bits isn’t necessarily enough for the temporary results.
It’s just a pet peeve of mine that DAC manufacturers ever say 32 bits - it’s too many for a PCM input and it’s not enough for the intermediate results needed for upsampling or (great) sigma delta modulation.
FWIW, this is the marketing paper I was talking about, it largely has to do with recording/pro audio. He’s talking about 32 bit integer PCM as an output format. Obviously, it’s not a technical paper, and my understanding of this is admittedly sophomoric.
" We continued our experiments trying to understand if and when PCM can challenge DSD. While 384k brought the quality closer, it was not until we started toggling between 24 and 32-bit depth that we heard a major improvement. It was clear that in a clean digital chain with 130dB dynamic range, 24-bit, even dithered, is the bottleneck.
Further analysis of the architecture of modern A-DC and D-AC chipsets shows that 32-bit decimation output is the information subset of a very high performance quasi-DSD front end. This allows the capturing of another 48dB of additional detail depth. Going back to 24-bit decimation reduces detail. 32-bit is needed to clean up the modern high performance digital recording PCM chain. 32-bit at a minimum of 352.8kHz (DXD32) PCM would be needed to compete with DSD sound quality."
I have their ADC/DAC setup that can do this, but the DAW software that will record/playback 32 bit integer (Wavelab/Cubase) requires another annoying USB dongle that I don’t want, so I haven’t tested myself.
IMO they are missing the boat slightly - 352.8Hz is too low of a sample rate to represent DSD adequately. Tho early on some DSD editing systems used 32 bit floating point (note that’s not the same as 32 bit fixed point) many DSD masterers and editors could hear the difference when DSD was converted to 32bit floating point 352.8kHz and back. In this specific case there’s definitely a difference between 32 bit floating point PCM and fixed point PCM. [Edit] I’d still argue that it’s not the DACs in DAC chips in particular that need a 32 bit format, it’s the digital processing elsewhere in the system (including any processing in the ADC/DAC.) I never argued that wider samples weren’t useful when processing, but they aren’t needed at the digital to analog proper. You get better results using more than 32 bits when filtering and converting to/from DSD. You don’t need 32 bits right at the DAC (after it does any processing.)
Wouldn’t voluntarily choose to disagree with anything Ted says at this point.
Even disregarding him having a digital brain the size of Betelgeuse…
You love your DS Dac I assume?
We’re luckily on Windom…so…personally…for me? …just sayin’…?
I’d argue with the next Guy.
I agree. I’ve heard some of those DSD-352.8khz conversions and it really didn’t do the DSD justice. I think for Mytek, even 352.8khz is already a really hard sell to audio engineers who still mostly haven’t ventured above 24/96 (or 44.1khz).
Btw, granted that it is hard to discern people’s tone on the internet, I’m really not trying to be disrespectful, testy, or argumentative. Just curious about some of these more technical things. Sorry for briefly eating up the bandwidth.
I didn’t take the posts as being disrespectful,… I hope I didn’t come across as testy. After seeing their paper it was clear to me where the confusion came from.
Me neither … I enjoy reading these posts where people wish to deepen their understanding of the D/A conversion principles, and I am glad for Ted taking the time to provide thoughtful answers.
