Why do NOS dacs sound different to oversampling designs?

[DonH56]

One of the notions I've been pondering, but haven't yet experimented with, is that the reproduced signal from those NOS DACs without an sufficiently effective analog reconstruction filter lack the time and inter-channel phase resolution of sharply bandlimited reconstruction. This echos back to the once popular (although, incorrect wrt the sampling theorem) notion that discrete sampling misses time/phase changes in the original sampled signal which are finer than the sampling interval. However, I believe that may actually be true for non-bandlimited NOS playback. If I correctly recall, brickwall filtering enables reconstructed waveform time and phase resolution finer than that of the sample interval. The sharper the bandlimiting, the finer the time/phase resolution and visa-versa. That is just conjecture, I currently have no evidence that this is the mechanism behind the fuzy/soft character that some of us perceive from NOS.

For these systems, the ADC side usually limits the bandwidth, since an antialising filter is required before conversion.

Brickwall filtering does not provide any more information than any other filter at the output of a DAC, just less out-of-band content. In fact, steeper filters introduce more phase shift and frequency droop/ripple in-band, below the Nyquist frequency (one-half the sampling rate).

Glitches, clock spurs, and low-level nonlinearities could account for some of the audible differences, among other things. Chief among "other things" in my mind, NOS or delta-sigma designs, are the quality of the image filter and output buffer.

My 0.000001 cent (microcent) - Don

 

[John Kenny]

Unary & binary combination like this, Don?
It's the output stage of the Soekris DAC - which is a 24bit R2R DAC which uses sign magnitude schema
If I'm right in my understanding of the schematic (only one register output is shown - each leg connects to a register output which is fed signals from an FPGA) - the first 3 MSBs are unary & the rest are binary R2R?

I can kinda understand how using unary signals & resistor structure on MSBs reduces glitches - less current required (less noise) when MSBs switch but isn't R matching more crucial for low level LSB signals rather than MSB signals?

Edit: Sorry, I get it now - the matching requirements of the R2R ladder are relaxed by the use of the unary MSB section. In modern commercial DACs, 6 MSBs are the usual unary section, thus further relaxing the matching requirements of the Rs in the R2R ladder. I wonder what the tradeoffs are in this approach & why Soekris DAC doesn't use 6 unary Rs in it's design?

 

[DonH56]

Hi John,

That looks like a typical hybrid design, yes.

Yes, lower switching currents can cause smaller glitches and nonlinearities from the glitches themselves as well as fewer settling issues for the buffers after the DAC.

As for matching, to use extreme examples the LSB needs to match only to about 50 % to achieve N-bit accuracy. It is only switching the smallest unit the DAC can change. The MSB has to match to an LSB; when it switches, the change is effectively half full scale, and right at the switching point you do not want more than an LSB change. So, the LSB needs to match 1 part in 2 (50 %), while the MSB of a 16-bit DAC needs to natch to within w part in 65,536 (0.0015 %, more or less). It's a bit more subtle than that but that is a hand-waving explanation.

If you split the MSB into parts, like in your picture, then the matching requirement is reduced (or error can be larger) by the number of parts. Much easier to match (usually with trimming) and now the midscale transition is smaller. Instead of switching every single bit, only one of the MSBs changes.

There are a lot of other design considerations in choosing the split (and DAC architecture in general) but hopefully that helps with the unary/binary trade.

HTH - Don

 

[John Kenny]

Thanks Don - I found again your excellent thread on WBF about this very topic & am having another read of it.
Takes a number of passes for this type of info to sink in & become absorbed

 

[DonH56]

Thanks for the kind words!

I missed your edit, sorry. The trade between the number of unary and binary bits gets complicated. Smaller incremental switching glitches and reduced matching requirements are offset by greater complexity (more power, area, sensitivity to process and thermal gradients, etc.), greater output capacitance (reduces bandwidth, extends settling time), more complex trim algorithms, etc. etc. etc.

 

[Ken Newton]

Brickwall filtering does not provide any more information than any other filter at the output of a DAC, just less out-of-band content. In fact, steeper filters introduce more phase shift and frequency droop/ripple in-band, below the Nyquist frequency (one-half the sampling rate)....

My 0.000001 cent (microcent) - Don

Hi, Don.

I'm not suggesting that brickwall filtering adds any new information, but rather, that it more accurately reveals the information that is there. Said another way, an absence of brickwall filtering will, according to the sampling throrem, prevent the original waveform from being accurately reconstructed. Band-limiting is an rather explicit requirement of the sampling theorem, and is necessary to accurately reconstruct the original continuous time signal. The D/A conversion process produces it's own set of image spectra having noting to do with the effectiveness of the A/D anti-alias filter. D/A converter output exists as discrete steps until/unless the image spectra are removed, thereby rendering the final output as continuous.

Steep slope SINC function FIR digital filters are easily made to exhibit linear phase simply by utilizing symmetrical coefficients in the filter kernel, as most do. Which, of course, is what produces the now infamous pre and post ringing impulse response of such filters, but their phase response is linear. As far as in-band response droop is concerned, even the technically compromised ubiquitous half-band FIR reconstruction filter (the kind incorporated in most DAC chips) are flat to within a very small fraction of a dB, including a zero order hold EQ function, fully up to 20KHz.

All of that said, despite it's near objective perfection, I'm usually most disappointed by the subjective sound of playback utilizing brickwall digital filtering. While I've yet to read a fully convincing technical argument for why that should be, I do believe we will eventually find the key technical parameter or factor that is adversely affecting human pshycho-acoustical perception of sharply band-limited audio sampling and reconstruction, despite the acknowledged mathematical perfection of the sampling theorem.

My nanocent.

 

[DonH56]

Hi Ken,

I think I may have misunderstood this whole thread, sorry. I was thinking of the output filter after a DAC, not a digital filter prior, or the filter in a delta-sigma loop (technically also prior to the DAC's output). I do not see how brickwall filter would be more revealing. Steeper filters can introduce other artifacts as you describe, and of course "softer" filters with larger transition regions allow more out-of band signal to pass through the system. One thing I have seen is out-of-band signals that get modulated back into the audible range through things like speaker modulation or mixing elsewhere in the chain. I do not know if the ear/brain does something similar or just quits responding over a certain frequency. Perhaps the brickwall filter's suppression of those higher signals and noise contribute to a higher sense of clarity in the sound, if not directly, indirectly by reducing artifacts (distortion) in the output buffer and rest of the chain to your ears. But that same lack may prevent signal from "filling in" lower frequencies that provide a sense of "richness" to the sound. Lots of examples of that sort of thing happening, and I know in the 1980's a lot of research was devoted to the digital noise floor and why adding noise created a more pleasing sound. It was not just reducing nonlinearities.

My guess, probably down to a picocent now... - Don

p.s. In my world of pulse-processing systems sinc^n functions were not used all that often. Of course, other filters may not provide the best response either... I did use sinc^n functions in several (very high speed, not audio) delta-sigma converters (ADCs and DACs) as they are fairly easy to implement in high-speed circuits.

 

[John Kenny]

I doubt it has anything to do with "acknowledged mathematical perfection of the sampling theorem." but rather it's attempted realisation in limited silicon estate.
On another thread we have the excellent analysis of 72 SRC software resampling

Just my femtocent

 

[DonH56]

An interesting question is whether or not we can really perceive the sampling interval in physical, say spatial, terms. At 44 kS/s we have one sample every 22.73 us, more or less. Assuming dry air at sea level the velocity of sound is about 1127 ft/s. That means sound travels about 0.31 inches in one sampling period. If you move your speaker 0.3", can you tell? Of course you'd have to level-match, and the real question is could you tell given a random relationship between the sampling point and signal (music). That seems to me a harder task than just listening for a shift in speaker position. Not sure people can detect that small a time interval? I do not know, but I think that relates to the question of whether or not we can hear "sampling time".

A side question is, if you just move your speaker 0.3" and do not do anything else, and you can tell, is it due to the change in loudness, change in position, or change in interaction with the room (frequency response)?

We're down to what, attocent? Which begs the question why zepto is next then yocto, putting z before y... I have had to deal with "atto" things but nothing smaller in my career so far. And no idea what's smaller than "yocto". So I've probably said all I can...

 

[Opus111]

I've been reading up (again) on DAC history for my newly started thread - Sony's very first CD player (CDP-101) did indeed have an 11uS time difference between the channels due to multiplexing a single DAC between L and R. I think it must have been a NOS player - the very first, long before 'NOS' became a buzzword. Stereophile reviewed it but they didn't have much software to play on it, you can read about it here : http://www.stereophile.com/cdplayers/193/index.html#f49v75ElgSRglvV8.97

 

[DonH56]

Yes, and most early DACs were 14 bits or so, and ADCs were pretty nonlinear. All the early CD players used conventional DACs AFAIK. It took a bit for digital processing technology to evolve and make delta-sigma designs more realizable. The first "single-bit" designs had some issues that were resolved as the architecture was better understood and technology and circuit design advanced. Dr. Gabor Temes was one of my grad profs, great guy!

 

[Ken Newton]

An interesting question is whether or not we can really perceive the sampling interval in physical, say spatial, terms....If you move your speaker 0.3", can you tell? Of course you'd have to level-match, and the real question is could you tell given a random relationship between the sampling point and signal (music). That seems to me a harder task than just listening for a shift in speaker position. Not sure people can detect that small a time interval? I do not know, but I think that relates to the question of whether or not we can hear "sampling time".

A side question is, if you just move your speaker 0.3" and do not do anything else, and you can tell, is it due to the change in loudness, change in position, or change in interaction with the room (frequency response)?

Yes, I agree, that is an interesting question. I would perhaps describe the main area of concern a bit differently. My concern would be with dynamic (program content generated) differences in transient timing and phase, especially between the two channels, not with fixed differences. A fixed time delay equates to an linear phase shift. Microsecond magnitude fixed inter-channel delays are not audible in my experience. However, dynamic phase and transient timing differences between channels are what produce the stereo illusion. Even so, the question of whether dynamic phase and timing differences produce a perceptable affect when they fall beneath a 22uS sample interval remains a valid one.

 

[amirm]

I've been reading up (again) on DAC history for my newly started thread - Sony's very first CD player (CDP-101) did indeed have an 11uS time difference between the channels due to multiplexing a single DAC between L and R. I think it must have been a NOS player - the very first, long before 'NOS' became a buzzword. Stereophile reviewed it but they didn't have much software to play on it, you can read about it here : http://www.stereophile.com/cdplayers/193/index.html#f49v75ElgSRglvV8.97

My first generation Technics (Panasonic) player also had a single DAC that was switched between channels.

 

[Ken Newton]

An interesting question is whether or not we can really perceive the sampling interval in physical, say spatial, terms. At 44 kS/s we have one sample every 22.73 us, more or less....

Don,

I just now stumbled across the below AES paper by Bob Stuart, et al. published in 2014. According to the paper, psychoacoustic research has shown the human ear can detect inter-channel transient timing differences as short as 10uS, and under certain conditions, down to 6uS! Intriguingly, the paper indicates that the sharper the filter slope the worse the distortion of transient timing, rather than the reverse which I thought was indicated by the sampling theorem. The logical conclusion of this notion is, NOS! Intriguing, indeed.

I believe the paper is intended as a foundational document to Meridian's MQA technology.

http://www.aes.org/tmpFiles/elib/20160408/17501.pdf

 

[Don Hills]

... psychoacoustic research has shown the human ear can detect inter-channel transient timing differences as short as 10uS, and under certain conditions, down to 6uS! Intriguingly, the paper indicates that the sharper the filter slope the worse the distortion of transient timing, rather than the reverse which I thought was indicated by the sampling theorem. ...

The timing resolution of 16/44.1 is more than adequate for sub-uS resolution. For a 0dBFS signal, it's in the region of 55 picoseconds.

Bob's sort of right about the relationship between filter slope and transient timing. The relationship between frequency resolution and time resolution is a see-saw - make one resolution finer, the other gets less well defined.

 

[DonH56]

I could not read the paper (got a "not found" error); may be because I let my AES membership lapse years ago. But I think I can bable a bit on your points anyway (when can I not? )

In general higher slopes/smaller transition band in the frequency domain leads to more ringing/aberrations in the time domain but filter theory is a complex field. It depends on the type of filter and its implementation. I have had a number of graduate-level classes but a lot of practical experience and it can still be challenging to choose and design the right filter for a given application.

I am not sure how filter slope ties to sampling theory? The theorem just says you must sample at twice the bandwidth of highest signal (information) frequency component. Note it is bandwidth, not absolute; if you put a bandpass filter around a 20 kHz signal at 1 MHz the theorem applies and you can sub-sample at 40.1 kHz and recover all the information in that 20 kHz band up at 1 MHz. If you have enough front end bandwidth and a plethora of other things that work....

NOS is not really my conclusion, because NOS requires a much sharper filter to prevent aliasing at the ADC input or to suppress images at the DAC's output. Oversampling allows you to use much softer filter slopes. You'll have more noise but less ringing. Another trade.

The aperture time for a 16-bit, 20 kHz signal is about 100 ps. Note aperture time, and thus jitter, is related only the signal frequency and resolution. The sampling frequency falls out of the equation and does not matter. More explanation over on the WBF articles. The 55 ps Don Hills (hi Don!) correctly asserted comes from assuming resolution of 1/2 lsb in either direction from the sampling point.

 

[Ken Newton]

I could not read the paper (got a "not found" error); may be because I let my AES membership lapse years ago. But I think I can bable a bit on your points anyway (when can I not? )

I'm now getting that same message. The link did work for awhile yesterday. At any rate, the paper is titled; "Convention Paper 9178. A Hierarchical Approach to Archiving and Distribution".

I am not sure how filter slope ties to sampling theory? The theorem just says you must sample at twice the bandwidth of highest signal (information) frequency component...

The concern over filter slope has to do with handling signals in a channel where the baseband signal bandwidth is close to the Nyquist frequency, which is the case with CD, where there is only 2kHz of guard band between the two. That, in conjunction with CD's 96dB dynamic range, dictate a brickwall filter slope to ensure that any alias products are suppressed to below the quantization noise floor. The sharper the brickwall filter the more suppressed will be any alias products. Signal channels where there is the luxury of a wider guard band can make use of more relaxed filtering and still suppress any alias products below the quantization noise floor. This appears to be approach taken by MQA.

The benefit of brickwall reconstruction filtering for CD audio playback is not so cut-and-dried wrt the sampling theorem. While the available guard band for suppressing playback image products during is just as narrow as it is for supressing recording alias products, unsupressed image products can not fold into the audio band as unsuppressed alias products can. So, one question becomes, if the image products are inaudibly ultrasonic is there any benefit to electronically filtering them - aside from protecting the following gain stage from misbehavior in the presence of fast slewing signals. Another question is whether supressing the ultrasonic image products plays any role in accurately reconstructing the timing of the original signal. Stuart's AES paper indicates that is not the case.

 

[amirm]

Your post Ken reminds of an AES presentation by FHG (makers of MP3) on a *lossy* codec that supported up to 96 Khz. As everyone knows, lossy codecs have a perceptual model of the human hearing system (http://www.audiosciencereview.com/forum/index.php?threads/fancy-terms-to-impress-your-friends.290/) and this codec had that too. At the end of the presentation by the FHG guy, a person raises his hand and asks, "if you can't hear anything above 20 Khz, how did you build your psychoacoustics model for region above 20 Khz?" The FHG guy froze in his tracks for good reason. If we can't hear what is up there, we can't test what works or doesn't for ultrasonics. In that sense, any random answer is as good as any other.

 

[Opus111]

So, one question becomes, if the image products are inaudibly ultrasonic is there any benefit to electronically filtering them - aside from protecting the following gain stage from misbehavior in the presence of fast slewing signals. Another question is whether supressing the ultrasonic image products plays any role in accurately reconstructing the timing of the original signal. Stuart's AES paper indicates that is not the case.

It seems to me, even if there's no slew limiting there may well be increased IMD in downstream circuits from not filtering out the image products. I recall a diagram from a Bruno Putzeys AES presentation indicating that removing the anti-imaging filter does have an impact on timing, I shall have a look to see if I can dig it out.

 

[DonH56]

I'm now getting that same message. The link did work for awhile yesterday. At any rate, the paper is titled; "Convention Paper 9178. A Hierarchical Approach to Archiving and Distribution".

The concern over filter slope has to do with handling signals in a channel where the baseband signal bandwidth is close to the Nyquist frequency, which is the case with CD, where there is only 2kHz of guard band between the two. That, in conjunction with CD's 96dB dynamic range, dictate a brickwall filter slope to ensure that any alias products are suppressed to below the quantization noise floor. The sharper the brickwall filter the more suppressed will be any alias products. Signal channels where there is the luxury of a wider guard band can make use of more relaxed filtering and still suppress any alias products below the quantization noise floor. This appears to be approach taken by MQA.

The benefit of brickwall reconstruction filtering for CD audio playback is not so cut-and-dried wrt the sampling theorem. While the available guard band for suppressing playback image products during is just as narrow as it is for supressing recording alias products, unsupressed image products can not fold into the audio band as unsuppressed alias products can. So, one question becomes, if the image products are inaudibly ultrasonic is there any benefit to electronically filtering them - aside from protecting the following gain stage from misbehavior in the presence of fast slewing signals. Another question is whether supressing the ultrasonic image products plays any role in accurately reconstructing the timing of the original signal. Stuart's AES paper indicates that is not the case.

My apologies, I was obviously speaking down to you earlier.

I certainly agree on narrow guard (transition) bands being a problem for filters. High-order filters are more subject to ringing, in-band amplitude and phase response is more problematic, and impedance variation is worse. Oversampling relaxes the filter order required for image suppression.

I agree with you and Opus111 that images can lead to other problems, including THD and IMD and stability problems with output buffers that are driven with out-of-band signals. There is also the possibility of overdriving tweeters and affecting the stability and performance of down-stream electronics (preamp, power amps).

I do not see how not suppressing images can affect timing except how high-frequency content impacts the circuits at and after the DAC's output stage. I do see how steep filters impact in-band phase, frequency and thus transient and time response.