|Notice one of the waveforms NOS.|
A few weeks ago on this blog in a comment discussion, Bennet / Dtmer Hk talked about showing what it looks like when upsampling a 1kHz 16/44.1 signal in PCM side-by-side with DSD upsampling.
Sure, no problem! We can have a quick peek at the 1kHz sine tone from a couple of DACs for comparisons between direct playback with built-in filtering, upsampled in high quality with PCM SoX, and then using high quality DSD conversion with SoX-DSD.
Plus, let's also have a look at the recent Beatles: Get Back documentary series and some thoughts which I think relate not just to the "music lover" which I hope is in all of us, but also to the "hardware audiophile" side of this hobby.
First, let's show the 1kHz sine upsampled in various ways as noted above. We start by defining the parameters I'll use in SoX 14.4.2. Starting with a clean 16/44.1, -0.2dBFS (a little overhead so as not to clip), 1.00227kHz (reduce rounding off at 44.1kHz) sine wave:
Converted To PCM 24/384 with:
sox input.wav -b 24 output.wav rate -v 384000 dither -S
These parameters were as recommended by Bennet and look like excellent choices if you want to do high quality upsampling. This will create a 24/384 (we could aim for 352.8kHz as well but honestly it makes no practical difference) output file with "very good" filtering. Also with high-pass dithering.
Converted also to DSD64/128/256 - using Måns Rullgård's SoX-DSD mods:
sox-dsd input.wav output.dsf rate -v 2822400 sdm -f sdm-8
<DSD rate = 2822400 for DSD64, 5644800 for DSD128, etc.>
These are the same parameters as discussed a few weeks back when testing DSD performance with the various DACs.
Using the TEAC UD-501 and Topping D10 Balanced DACs, captured with E1DA Cosmos ADC (mono mode), here are a few side-by-side high resolution FFTs - captured using FlexASIO at 32/384 (192 bandwidth), 512k-points:
As you can see, the TEAC graph includes the NOS mode where we can see all the ultrasonic imaging artifacts when we run the DAC filterless, top right is the standard direct playback with the typical "sharp" filter I use the vast majority of the time. The more modern Topping D10 Balanced doesn't have a NOS mode but can handle DSD256, pushing the ultrasonic noise shaping further out. As shown back in August, the D10 Balanced has by default a steep minimum phase filter for direct playback.
If we do some pixel peeping of the SoX PCM upsampled FFTs, we can make out the steeper and deeper post-Nyquist attenuation compared to the DAC's direct playback thanks to SoX's "very" high quality sharp linear filter. Of course, DACs do have noise floor limits so there's only so much attenuation possible. Even if the low-pass filter applied is capable of mathematically achieving 32-bits with better than -190dB attenuation, that's just a numbers game (that sometimes gets tossed around as if meaningful) and in real life we're not going to see that magnitude in actual playback from the analogue outputs.
Note that distortion measurements like the THD / THD+N do not and really should not change much** with various PCM and DSD processing/upsampling. Feeding a DAC with 24/384 data compared to 16/44.1 over USB could account for some of the change and likewise DSD vs. PCM will also lead to variation. Do not assume that big numbers like 384kHz upsampling or DSD256 will automatically lead to higher fidelity - as you can see, this is not necessarily the case. This is also why having test data for your specific DAC is essential.
[** Interestingly, the biggest difference I see is with the Topping D10B going from 16/44.1 to 24/384 PCM. Notice the increase in THD. I wonder if we're seeing the same phenomenon as the Topping D90SE at and above 192kHz suggesting that there are potential optimizations that can be tweaked with the ESS THD compensation parameters at higher samplerates.]
Like with the measurements, I don't hear a difference in the sound between various filters for the most part although there are exceptions. For example, the slow roll-off filter with the Ayre/PonoPlayer a few years ago is an example of one that did change the sound noticeable. Sure, we could do fancy stuff to allow even more precise filtering - Chord's WTA, HQPlayer upsample, even try out PGGB if you want. You can also play with minimum or intermediate phase settings (as discussed previously).
What's important IMO is that we keep the effect of upsampling in context and realize that this is just low-pass filtering, not more-complex DSP like FIR room correction. As per this post from Måns, you don't need "megatap/megaorder" filters to achieve more than adequate very low passband ripple and large stopband rejection. As he showed, ~400-taps is more than enough, providing 180dB stopband attenuation for any amplifier, and certainly for any human ears. Make sure to let that sink in so that we can be better informed consumers.
Again, big numbers might seem impressive at first and perhaps can spur sales based on specs, but in reality, this is just not needed and eventually people get wise to the marketing and become unhappy if pushed beyond the realm of believability. Looking at the SoX PCM upsampling debug log, it's "only" using a 573-taps FIR with 186dB stopband attenuation for the 44.1 upsampling to 32/384kHz internally. This can be increased to 645 taps with 210dB attenuation if we use -u instead of -v for "ultra" quality.
As usual, for those who advocate for really really complex upsampling algorithms and enormous tap lengths, it would be nice to see evidence that this makes a difference with actual DAC output rather than supposed "subjective" claims of improvement with no apparent basis in physical reality. Even if one insists that ears are the best tools to appreciate differences because one doesn't believe that measurements can capture supposedly audible changes, let's get some controlled listening (like blind testing) to demonstrate/confirm the claims!
Be mindful of these ideas when you're on audiophile forums and people are talking about using stuff like powerful Intel i9, Xeon, AMD Ryzen 9, Threadripper CPUs, and leveraging the power of GPUs to perform all their upsampling whether to PCM 768+kHz or DSD512+ as if this too is some kind of impressive feat. I think we've all seen examples of multi-thousand-dollar computers being used for this purpose with all kinds of claims that it improves sound (like this product). IMO, this is all highly unlikely unless one is also applying room-correction or other type of DSP that can definitively change the sound.
Audiophiles spend a lot of money (and time) on things which I think are unnecessary and suspicious, with claims of subjective improvements. Hey, if this is part of one's "pursuit of happiness", sure, go have fun! Just make sure in principle not to entertain too much snake-oil or be too giddy about fantasies, because at the very least, that would be bad for the reputation of the audiophile hobby as a reality-based technical endeavor.
|Starr & Preston playing with the Stylophone. Lennon taking a sip.|