What do you think makes NOS sound different?

[Ken Newton]

Some of us prefer the sound of non-oversampled(NOS) DACs, while others of us prefer oversampled. However you may feel about the difference you may hear in subjective sound character between the two, it's my sense is that most audiophiles can readily observe that there is some kind of difference audible between oversampled(OS) and NOS DACs. The overriding subjective characteristic which I hear from NOS is an immediate and obvious sense of psychological ease. While early OS DACs too often sounded annoying, or, what’s almost as bad, sounded boring, I find that there can still be an subjective advantage for NOS in terms of, dare I say, a vinyl-like musical ease. I don’t readily notice that such ease is lacking from modern OS DACs - that is, until I switch to a NOS DAC. Not everything, however, subjectively favors NOS over OS. For example, to my ears, OS presents a much wider soundstage, albeit not as deep nor as 3D as NOS presents. In addition, NOS bass tends to sound relatively ‘unfocused’, for lack of a better term. The bass is fully present in energy, just not as punchy or coherent sounding.

There are well understood reasons for the OBJECTIVE technical differences. Prime among these are the repeating image-bands, ALL of which are ultrasonic, so, are not only inaudible to the human ear, but are not reproducible by most loudspeaker tweeters anyhow. The lowest image-frequency enabled by the CD standard begins in the ultrasonic, at 22KHz, but CD standard recording anti-alias filters usually only maintain flat signal response up to 20KHz, which means that the lowest flat (non-rolled off) image-frequency is: 44.1KHz - 20KHz = 24.1KHz, which is even further in to the ultrasonic.

Another technical difference is the famous NOS band-edge response droop of -3.16dB @ 20KHz, due to the SINC aperture inherently produced by Zeroth Order Hold (sample then hold) NOS multibit DAC operation. However, this droop is also largely near the ultrasonic. Also, while it does rapidly decrease in effect with decreasing frequency, it should be mentioned that uncorrected, the droop’s effect extends down to about 5KHz. However, the droop is easily corrected via simple analog EQ without harming the NOS character. There are other less gross technical effects as well, such as a decreased in-band quantization noise floor for oversampling. Also, some differing jitter character between OS and NOS.

I want to make explicitly clear, that I’m not expressing any set preference for NOS over OS. The issue here is not that one may sound better than the other, but rather, why do they sound different at all? A particularly perplexing matter with signals that fall well short of the upper audible band-edge. Such as it does for the human voice. The lower the baseband signal, the higher will be it's image frequency. So, we’re talking of image frequencies far in to the ultrasonic for vocals. I want to emphasize this particular point about the audible difference between NOS vs. OS persisting even with lower frequency range content which have no significant overtones near the OS filter cut-off. Maybe there are intermodulation effects being produced by some stage(s) in the amplification chain in response to any unsuppressed image-bands, but the NOS character persists even when passively band limiting the DAC chip’s ultrasonic output to prevent that.

The only explanation I can readily come up with that fits all of the observations above is, inadequate implementation of an OS DAC’s digital interpolation filter, and which was long ago identified by Lagadec as producing time-domain artifacts not present in sampling theory. Equiripple interpolation filter implementations. Which are, essentially, ubiquitous among DAC designs, This is evidenced by the seemingly inconsequential and uniformly repeating frequency response ripples across an equiripple filter’s pass band. These ripples are visible within most OS DAC chip’s integrated digital filter data-sheet performance charts. If that is the reason for the subjective difference between OS and NOS, perhaps NOS is closer to how OS DACs SHOULD sound, and would sound without certain, typically chip implemented, OS digital filter processing artifacts. That's only speculation.

What is your thinking on the technical reason for the subjective difference, even for lower range sounds, between OS and NOS DACs?


HELPFUL THREAD LINKS: ---===========================================================================================
1) Initial Suspect List. Post 153: https://www.diyaudio.com/community/...os-sound-different.371931/page-8#post-6654353

2) Reconstruction/Image-Band Handling. Post 344: https://www.diyaudio.com/community/...s-sound-different.371931/page-18#post-6668363

3) Comprehensive Results Document for 44.1kHz, 88.2kHz and 176.4kHz Resampling Experiments. Post 1773: https://www.diyaudio.com/community/...s-sound-different.371931/page-89#post-6775491

4) Suspect List, with item assessments added. Post 1149: https://www.diyaudio.com/community/...s-sound-different.371931/page-58#post-6734286

5) Implicated Suspect Category. Post 1192: https://www.diyaudio.com/community/...s-sound-different.371931/page-60#post-6736142

6) Filter Signal Echo/Reflection Experiment Introduction, and Results Analysis. Posts 1043, and 1518: https://www.diyaudio.com/community/...s-sound-different.371931/page-53#post-6727251
https://www.diyaudio.com/community/...s-sound-different.371931/page-76#post-6761338

7) Investigation Concluding Report. Post 1783: https://www.diyaudio.com/community/...makes-nos-sound-different.371931/post-6777645

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

8) Frans Sessink’s Settling-Time Error Mitigation Presentation. Post 229: https://www.diyaudio.com/community/...s-sound-different.371931/page-12#post-6658063

9) Marcel van de Gevel’s Dither paper. Post 1720: https://www.diyaudio.com/community/...s-sound-different.371931/page-86#post-6770852

10) Improved, ‘Sonic Scrambling’ Subtractive Dither Concept, Developed by Marcel van de Gevel and Hans Polak. Post 1692 and 1756: https://www.diyaudio.com/community/...s-sound-different.371931/page-85#post-6769819
https://www.diyaudio.com/community/...s-sound-different.371931/page-88#post-6773204

• Remastero's PGGB-RT: A Free, High-Performance Interpolation-Filter plug-in for Foobar2K. Software-Tools sub-forum. PGGB upsampling software
• Announcement of new and improved higher precision 64-bit version of PGGB-RT Foobar2000 plug-in module. https://www.diyaudio.com/community/threads/pggb-upsampling-software.376376/page-2#post-7316157

 

[rfbrw]

Are you sure you want to open this can of worms ? The definition of the term NOS alone is probably a thread of its own.

 

[Ken Newton]

Yep, I sure do. I hope that someone has a logical technical explanation.

As for the acronym, NOS. While it may be a bit in-artful, I think its commonly understood to signify a DAC without an FIR interpolation filter. Leaving the D/A output as an un-reconstructed signal. Either retaining it's image bands, or somewhat suppressing them post D/A conversion in the analog domain.

 

[rfbrw]

I don't think the missing filter is it. The Sony CDP-101 is NOS but it is by far the worst player I have ever heard. NOS PCM63's, OTOH, have their moments.

 

[Don Hills]

A reconstruction filter is mandatory for compliance with the sampling theorem. Without one, you get the problems you have already described. In addition, transient timing goes all to heck - it deteriorates from as low as 110 picoseconds to as much as 22 microseconds. There are several studies claiming that microsecond timing errors are audible. So you may well hear a difference.

 

[abraxalito]

Its a very interesting topic - based on my experiments of adding upsampling to my NOS DACs over many years, the non-fatiguing nature of NOS doesn't evaporate when feeding them at 88.2kHz. So my hypothesis its not NOS per se that makes the difference to the fatigue factor, rather its some correlate of typical NOS designs that's doing it.

 

[Ken Newton]

I don't think the missing filter is it. The Sony CDP-101 is NOS but it is by far the worst player I have ever heard. NOS PCM63's, OTOH, have their moments.

One of the main problems in assessing what is responsible for the poor sound of a given component is the multiplicity of sub-system elements it may contain. So, it becomes impossible to know whether it's the digital filter, or quantizer error, or clock jitter, or I/V conversion, or analog output stage AC errors, or supply regulation, or what have you, that is at at fault. Or, perhaps, it's some combination of such elements in a given product. One must have a highly modular DAC, it would seem, in order to conduct a controlled search to identify the culprit of a poor sound quality.

 

[Ken Newton]

A reconstruction filter is mandatory for compliance with the sampling theorem. Without one, you get the problems you have already described. In addition, transient timing goes all to heck - it deteriorates from as low as 110 picoseconds to as much as 22 microseconds. There are several studies claiming that microsecond timing errors are audible. So you may well hear a difference.

Hi, Don,

Except that, with digital audio, there are inherently two stages of image supression which are always applied. The listeners ears, and the loudspeakers tweeters, since all of the image bands are ultrasonic.

I'll have to give more thought to your intriguing observation about transient timing. I once saw in a presentation, I think it was by (Bruno Putzys), NOS features a time variant impulse-response. Which is what I believe you are saying. Now, why that should make NOS sound more 'analog' and natural, I have no idea.

One question, though. That 22uS maximum figure is obviously the 44.1KHz sample interval, however, from where does the 110pS minimum figure derive?

 

[Ken Newton]

Its a very interesting topic - based on my experiments of adding upsampling to my NOS DACs over many years, the non-fatiguing nature of NOS doesn't evaporate when feeding them at 88.2kHz. So my hypothesis its not NOS per se that makes the difference to the fatigue factor, rather its some correlate of typical NOS designs that's doing it.

Hi, Richard,

Yes, I've followed some of your reports on your encouraging experiments with upsampling via the SoX utility within Foobar. One of my current pet theories is that all of the best sounding upsamplers utilize a windowed SINC FIR kernel. I believe this was even true about the once very popular PMD-100 filter chip, and now, the SoX utility. Most silicon based oversampled interpolation filters are Parks-McClellen half-band. All of which seem to exhibit the very tiny and uniformly sized, frequency domain passband response ripples indicative of time-domain signal reflection artifacts within the filter. But that suggested correlation is just speculation on my part for now.

I haven't really searched, but since you've been working with Foobar, do you know of a small and inexpensive computing module to which the SoX utility or, perhaps, Foobar itself, may be ported? With the objective of possibly embedding it within a diy DAC box and inserted between the DIR and the DAC functional blocks via I2S.

 

[Don Hills]

...
One question, though. That 22uS maximum figure is obviously the 44.1KHz sample interval, however, from where does the 110pS minimum figure derive?

Here's something I prepared earlier...
The time resolution of a 16 bit, 44.1khz PCM channel is not limited to the 22.7µs time difference between samples. The actual minimum time resolution is equivalent to 1/(pi * quantization levels * sample rate). For 16/44.1, that is 1/(pi * 65536 * 44100), which is about 110 picoseconds. To put that in perspective, light travels a bit over an inch in that time.

Shannon and Nyquist showed that as long as you keep all components of the input signal below half the sampling frequency, you can reconstruct the original signal perfectly - not just in terms of amplitude, but in terms of temporal relationships too. They only addressed sampling, and assumed infinite resolution in amplitude. With a digital signal the precision is limited by the number of amplitude steps, leading to the above formula.

If you want to see a real world demonstration of a single event (the edge of a square wave) being accurately sampled between sample points, check out Monty's show and tell video at the 20:55 mark. If anyone hasn't seen the video before, I strongly suggest you take the time to watch it all.

D/A and A/D | Digital Show and Tell (Monty Montgomery @ xiph.org) - YouTube

 

[Don Hills]

... One of my current pet theories is that all of the best sounding upsamplers utilize a windowed SINC FIR kernel...

The mathematically correct reconstruction filter is a SINC function.

For upsampling, take a look at the PGGB utility.
PGGB - Offline remastering

 

[abraxalito]

Yes, I've followed some of your reports on your encouraging experiments with upsampling via the SoX utility within Foobar. One of my current pet theories is that all of the best sounding upsamplers utilize a windowed SINC FIR kernel. I believe this was even true about the once very popular PMD-100 filter chip, and now, the SoX utility. Most silicon based oversampled interpolation filters are Parks-McClellen half-band. All of which seem to exhibit the very tiny and uniformly sized, frequency domain passband response ripples indicative of time-domain signal reflection artifacts within the filter. But that suggested correlation is just speculation on my part for now.

Hi Ken - yes I've been wondering if the half-band FIR filter was one major contributor to the fatiguing sound. Fortunately I'm fairly close now to being able to test this hypothesis as I have a prototype DAC with an on-board MCU with enough computing clout to run the kind of filters that used to be ubiquitous in DAC designs in the late 1980s and are embedded in such chips as PMD-100. I have already calculated a couple of example filters to try out, I just need code to run them now : Cheap ARM MCUs for RBCD audio

I haven't really searched, but since you've been working with Foobar, do you know of a small and inexpensive computing module to which the SoX utility or, perhaps, Foobar itself, may be ported? With the objective of possibly embedding it within a diy DAC box and inserted between the DIR and the DAC functional blocks via I2S.

That is indeed the kind of project I've been thinking about the past couple of weeks. I suspect that an MCU as the STM32H750 might well be able to do the trick for practical FIR filters up to a couple of hundred taps. This would only run the FIR kernel, no fancy bells and whistles.

 

[Extreme_Boky]

analog: you have to spin it / scratch it to hear it, right? nice and smooth spinning / smooth gliding - better sound

digital: you have to switch it to hear it, right? the less you switch it - the better (the more correct) it sounds. NOS - not much switching, keep the switching to a minimum (small amplitudes also help)

not many people are after the correctness. what is it anyway? to me, it is the closest I can get to ... an analog sound; the most natural sound I can get with digital reproduction.

many people like oversampling to insane rates PCM and/or DSD. it does sound impressive, but not natural. I have May DAC that can do crazy things, but by far the most enjoyable sound is at the native sampling rate, NOS mode. no oversampling (by HQ Player, for example, and no oversampling by May).

but, many will disagree...

 

[hop333]

Very interesting discussion. I am also very interested in correlating clear sound quality differences in the A/D and D/A chain with something that can be measured or reasoned about.

I'm not talking about people claiming that a certain resistor or op amp has a certain sonic signature which I doubt very much having given people the opportunity to hear very different components in controlled double blind ABX tests with levels and frequency responses set carefully many times. They inevitably fail to even identify the components let alone express a preference. One must never forget the effect of psychology in these tests, and try to remove it as far as possible. In contrast, small level differences across the band or in parts of the band can be heard consistently. One passive component that could be heard with a low but non-zero probability was the 220µF bipolar electrolytic capacitor on the input of a Quad ESL63 being replaced by a huge plastic film, by the way.

Since my early years working as a recording engineer and then as a loudspeaker designer, it has been clear to me that the biggest challenges to get right in audio are the transducers. I used to think of this as microphones at one end and loudspeakers at the other. I'm now almost ready to add A/D and D/A to that list in the sense of energy being converted between forms.

The idea of a CPU or GPU to explore different digital filters and other D/A mechanisms in real time is very appealing. Interesting filters that have not been discussed in this area to my knowledge which adapt to the signal (for example the autocovariance structure within a channel and between channels) are one possibility. Of course ABX testing would also be a piece of cake compared to the randomised box with top quality relays etc that were required for testing physical components.

 

[lcsaszar]

Another technical difference is the famous NOS band-edge response droop of -3.16dB @ 20KHz, due to the SINC aperture inherently produced by zeroth order hold (sample then hold) NOS multibit DAC operation. However, this droop is also mostly near ultrasonic. Also, while it does rapidly decrease in effect with decreasing frequency, it should be mentioned that uncorrected, the droop’s affect extends down to about 5KHz. However, this is easily corrected via simple analog EQ without harm to the NOS character.

Zero-order hold is just one implementation. I am not very good in math, but I think a very short hold time and zero between samples would give a flat response in the passband. The average energy will be lower, and there will be more harmonics of the sampling frequency beyond the audio band. The idea is to mimic the sampling in the ADC which is instantenous, not S/H like.

 

[Ken Newton]

The mathematically correct reconstruction filter is a SINC function.

For upsampling, take a look at the PGGB utility.
PGGB - Offline remastering

Yes, it is. The problem is that some implementations of FIR band-limiting filters produce processing artifacts and problems. One of the biggest offenders is how half-band filters just plainly violates the sampling theorem by not sufficiently band-limiting the signal within the constraints of the CD specification, due to the insufficiently wide 2KHz guard-band available between 20KHz and 22KHz. Until proven otherwise, my suspicion is directed at the ubiquitous equiripple FIR interpolation filters.

Anyhow, for the benefit of those who may not be familiar, here is a link to a paper by the late Julian Dunn, written more than two decades ago, discussing the time-domain issue with equiripple FIR filters that Lagadec had first identified.

https://www.nanophon.com/audio/antialia.pdf

 

[Ken Newton]

That is indeed the kind of project I've been thinking about the past couple of weeks. I suspect that an MCU as the STM32H750 might well be able to do the trick for practical FIR filters up to a couple of hundred taps. This would only run the FIR kernel, no fancy bells and whistles.

1) This sounds very interesting. Please keep us apprised of your developments on this.

2) I'm curious, and I've probably missed your conclusions on this, but did you find the Foobar based x2 upsampling of your NOS DAC to retain all the positive subjective attributes of NOS sound? Also, did it merge them with any of the positive subjective attributes which OS can exhibit? The best of both worlds, as it were.

 

[hop333]

My current investigation is why a well-engineered first generation BitStream DAC (Meridian 203 with 2x SAA7321GP run in dual mono differential mode and attention paid to jitter) with apparently good analogue & PSU design and which measures very well, sounds so poor. For example, the highest frequencies like cymbals and their harmonics sound so synthetic, nothing like metal that is ringing.

I guess it's possible that there is something wrong with it, or that a/some critical electrolytic caps have failed - after all it is about 30 years old now. But I don't remember it even sounding that great in the early 90s when I first got it.

I am currently looking at differential analogue stages (better op amps, high-quality audio output transformers, getting rid of the DC servo) but these may be red herrings. Perhaps something about the raw voltages coming out of the DAC is the problem, and if so what? Having said that, transformer out has apparently made cymbals sound more realistic so at least some of this might be down to analogue engineering.

 

[Ken Newton]

Since my early years working as a recording engineer and then as a loudspeaker designer, it has been clear to me that the biggest challenges to get right in audio are the transducers. I used to think of this as microphones at one end and loudspeakers at the other. I'm now almost ready to add A/D and D/A to that list in the sense of energy being converted between forms.

I concur, both the loudspeaker AND it's interface with the room are responsible for many subjective sins which are often and incorrectly attributed to the source format or to the electronics.

 

[Ken Newton]

I am currently looking at differential analogue stages (better op amps, high-quality audio output transformers, getting rid of the DC servo) but these may be red herrings. Perhaps something about the raw voltages coming out of the DAC is the problem, and if so what? Having said that, transformer out has apparently made cymbals sound more realistic so at least some of this might be down to analogue engineering.

Sounds like you are on the right path. One of the things about DACs which attracts me and others as hobbyists, is the same thing which seems to make them so difficult to fully analyze. Their multiple complex sub-systems and mixed-signal nature.