When I first heard of NOS-DAC's some years ago, I shaked my head and thought the whole idea was totally nuts. If you look at a 15khz sine wave from a NOS-DAC, it looks like some sort of disease. The wave is jagged and should sound sharp and terrible.
But since then I have pondered the subject and nowadays I'm a bit of a believer. The turnaround came when I began distinguishing between amplitude related and time related distortion.
For me, it makes sense that our hearing is particularly sensitive to time related distortion. Many things supports this. First of all, our hearing makes judgments of the various sounds from the order which they appear, with reflections and such things. Secondly, it seems that many audiophile “myths” supports this. Many think negative feedback in an amplifier degrades the sound. They prefer perhaps a single ended class A tube amp before some scientifically crafted class AB monster.
Well, the thing is that the former mainly produces amplitude related distortion and the latter mainly time related. The NFB amp will distort much less, but the feedback will inevitably introduce a “smearing” effect at high frequencies. Actually a bit like the oversampling DAC's.
Many think an oversampling DAC sort of “fills in” the missing information. That's wrong, you can't recreate lost information ( mostly , actually Naims new format-MQA- is said to do just that ). What the OS does is to create a “pendulum effect” that is supposed to smooth out the jagged shape of , let's say, a 15khz signal. In order to do so, the interpolator – a FIR filter – looks backwards and “recreates” the sine wave. You can say this algorithm has an averaging effect. And this is probably the culprit.
In order to make sine waves look good at the oscilloscope ( very good amplitude related distortion) it assumes that this very wave doesn't change it's amplitude abruptly. If so, we have that famous smearing effect. The OS DAC trades amplitude related distortion with time related.
Look at the attached pictures from my oscilloscope. It's a 15khz sine wave in both NOS mode and oversampled mode. It's from my PCM1704 DAC and the rounded shape of the NOS output comes from the fact that I'm using a 3rd order analog Butterworth filter. I think it rounds off things nicely – look at the diagram of the filter. It should spare the following electronics from too much noise.
OK, compare it with the same DAC but with the interpolation filter enabled. It “smears” out things and makes them look better!
On the other hand, NOS-DAC's produces a lot of overtones that may be both audible and also disturbing to the following circuits. But - this distortion isn't really time related, is it! But those artifacts will probably be heard, and that's why it's always a matter of taste if you like NOS or not.
My own personal impressions are that a NOS DAC sounds both more round and soft and the treble is like flaky cotton. Summed together, a NOS-DAC sounds a bit like a juicy fruit cake.
The oversampling version sounds more “correct” and also a bit thin and sharp. There is the usual problem with “s” sounds and strings.
In the future, when the internet will have much more capacity, all audio will probably be transferred in 196/24 format, and then all these concerns will be only a memory. With such a high fs all DAC's may be NOS.
The sigma delta DAC's are also considered to be slightly inferior to good multibit ones. Also here we have a lack of time related accuracy. The S/D DAC relies on an averaging effect and can never present those minute rapid changes of the signal that a multibit DAC can. Mathematically, S/D DAC's are more or less flawless. But when the signal changes rapidly it loses precision.
But if we take the PCM1794 DAC as an example, it should sound very good. It has a 6 (or 7 in mono mode ) bit hardware ladder that is combined with a 512Fs ( 8 times oversampled ) S/D algorithm, we actually have 65535 discrete levels ( 512fs * 2^(7 bits) ). And this is the same as a 16 bit multibit DAC.
Finally, I think the bad reputation of OS DAC's comes from the aggressive reconstruction interpolator. But on the other hand, those totally unfiltered NOS DAC's out there are an extreme at the opposite direction.
But since then I have pondered the subject and nowadays I'm a bit of a believer. The turnaround came when I began distinguishing between amplitude related and time related distortion.
For me, it makes sense that our hearing is particularly sensitive to time related distortion. Many things supports this. First of all, our hearing makes judgments of the various sounds from the order which they appear, with reflections and such things. Secondly, it seems that many audiophile “myths” supports this. Many think negative feedback in an amplifier degrades the sound. They prefer perhaps a single ended class A tube amp before some scientifically crafted class AB monster.
Well, the thing is that the former mainly produces amplitude related distortion and the latter mainly time related. The NFB amp will distort much less, but the feedback will inevitably introduce a “smearing” effect at high frequencies. Actually a bit like the oversampling DAC's.
Many think an oversampling DAC sort of “fills in” the missing information. That's wrong, you can't recreate lost information ( mostly , actually Naims new format-MQA- is said to do just that ). What the OS does is to create a “pendulum effect” that is supposed to smooth out the jagged shape of , let's say, a 15khz signal. In order to do so, the interpolator – a FIR filter – looks backwards and “recreates” the sine wave. You can say this algorithm has an averaging effect. And this is probably the culprit.
In order to make sine waves look good at the oscilloscope ( very good amplitude related distortion) it assumes that this very wave doesn't change it's amplitude abruptly. If so, we have that famous smearing effect. The OS DAC trades amplitude related distortion with time related.
Look at the attached pictures from my oscilloscope. It's a 15khz sine wave in both NOS mode and oversampled mode. It's from my PCM1704 DAC and the rounded shape of the NOS output comes from the fact that I'm using a 3rd order analog Butterworth filter. I think it rounds off things nicely – look at the diagram of the filter. It should spare the following electronics from too much noise.
OK, compare it with the same DAC but with the interpolation filter enabled. It “smears” out things and makes them look better!
On the other hand, NOS-DAC's produces a lot of overtones that may be both audible and also disturbing to the following circuits. But - this distortion isn't really time related, is it! But those artifacts will probably be heard, and that's why it's always a matter of taste if you like NOS or not.
My own personal impressions are that a NOS DAC sounds both more round and soft and the treble is like flaky cotton. Summed together, a NOS-DAC sounds a bit like a juicy fruit cake.
The oversampling version sounds more “correct” and also a bit thin and sharp. There is the usual problem with “s” sounds and strings.
In the future, when the internet will have much more capacity, all audio will probably be transferred in 196/24 format, and then all these concerns will be only a memory. With such a high fs all DAC's may be NOS.
The sigma delta DAC's are also considered to be slightly inferior to good multibit ones. Also here we have a lack of time related accuracy. The S/D DAC relies on an averaging effect and can never present those minute rapid changes of the signal that a multibit DAC can. Mathematically, S/D DAC's are more or less flawless. But when the signal changes rapidly it loses precision.
But if we take the PCM1794 DAC as an example, it should sound very good. It has a 6 (or 7 in mono mode ) bit hardware ladder that is combined with a 512Fs ( 8 times oversampled ) S/D algorithm, we actually have 65535 discrete levels ( 512fs * 2^(7 bits) ). And this is the same as a 16 bit multibit DAC.
Finally, I think the bad reputation of OS DAC's comes from the aggressive reconstruction interpolator. But on the other hand, those totally unfiltered NOS DAC's out there are an extreme at the opposite direction.
Attachments
You may be interested in AES paper by Rob Stuart (2014).
The sharper the filter slope the worse the distortion of transient timing. It's known that the attack phase of sounds is the most important phase of any sound.
I included the pdf. It is a must-read.
Last edited:
Thanks, I was aware that JRS was a fan of 'leaky' filters which allow substantial aliasing. Which then violates the Nyquist criterion and that, according to Bruno Putzeys means the timing accuracy is compromised. So which to choose - Nyquist violation on a grand scale or 'time smearing' ?
Just scanning the paper - why would he give an example of 8 30kHz 2nd order filters in sequence? The frequency response being shot at 20kHz from such would mean no self-respecting engineer would introduce that degree of HF attenuation in a line-level audio product. Completely unrealistic in my estimation.
Just scanning the paper - why would he give an example of 8 30kHz 2nd order filters in sequence? The frequency response being shot at 20kHz from such would mean no self-respecting engineer would introduce that degree of HF attenuation in a line-level audio product. Completely unrealistic in my estimation.
Last edited:
The waveforms of your NOS DAC look staircase-like, which means that there is a zeroth order hold somewhere. Maybe you just like the slight treble roll-off caused by this zeroth order hold.
With a true NOS design, the DAC chip will output the exact same sample rate as it sees it at the input of the whole DAC unit. There are no filters; digital nor analogue. Many people discard the true NOS design because it sounds different from other DAC’s, and it looks “ugly”.
I like the true NOS design because it is a minimalist approach to a digital to analogue conversion. The current-out (if sufficiently high in value) can be converted by the simplest means of a resistor, which keeps the minimalist approach intact.
The sample-and-hold you are referring is a normal behaviour of pure NOS design. The staircase sinewave should stay a true staircase waveform even at 20khz.
The treble roll-off is an inherent characteristic of all DAC’s (not only true NOS ones), because the 22kHz Nyquist frequency drops within an audible range.
The above issue is easily fixed by oversampling (all commercial DAC’s oversample).
Some people like to send an oversampled signal to a true NOS DAC.
I like the true NOS design because it is a minimalist approach to a digital to analogue conversion. The current-out (if sufficiently high in value) can be converted by the simplest means of a resistor, which keeps the minimalist approach intact.
The sample-and-hold you are referring is a normal behaviour of pure NOS design. The staircase sinewave should stay a true staircase waveform even at 20khz.
The treble roll-off is an inherent characteristic of all DAC’s (not only true NOS ones), because the 22kHz Nyquist frequency drops within an audible range.
The above issue is easily fixed by oversampling (all commercial DAC’s oversample).
Some people like to send an oversampled signal to a true NOS DAC.
I was under the impression that 'NOS' meant 'no oversampling' - hence I see no reason why a DAC which uses no oversampling but does use a reconstruction filter shouldn't be called 'NOS'. All the NOS DACs I've ever seen do use some form of filtering, normally just a 2n2 in parallel with the I/V resistor.
Indeed. Of course the reconstruction filter of a non-oversampling DAC should peak a bit to compensate for the zeroth order hold filter if you want a flat response; with a simple RC parallel network as filter the roll-off only gets worse.
Anyway, smooth treble roll-off can easily explain the thread starter's subjective impression, can't it?
Anyway, smooth treble roll-off can easily explain the thread starter's subjective impression, can't it?
Last edited:
The capacitance is always there. Not necessarily at the I/V resistor, but embedded in the interconnects, or at the input stage of an amp. This could be expanded further to include the speaker cable capacitance, tweeters' roll-off at high frequency, and ultimately the human ear natural roll-off characteristics.
(Now that I mentioned few above, it does make sense not to include any capacitance at I/V resistor... at least to me
)
The NOS does indeed mean no oversampling, but we need to come up with the name for a NOS DAC that does not use digital nor analog filtering ...
Is the traditional NOS design (that could include a filter at some point) still a NOS design if you feed it oversampled material?
(Now that I mentioned few above, it does make sense not to include any capacitance at I/V resistor... at least to me
)The NOS does indeed mean no oversampling, but we need to come up with the name for a NOS DAC that does not use digital nor analog filtering ...
Is the traditional NOS design (that could include a filter at some point) still a NOS design if you feed it oversampled material?
People who want a flat response (that is just a representation of the frequency response on an oscilloscope, by the way, with absolutely no time-domain correlation info of any kind..) should look at oversampling DAC's.
After many years of listening to various DAC's, I have firmly settled on a "true" NOS design as my favorite. I also have other DAC's that I listen to and tend to enjoy, but the one that lets me immerse myself into music for extended periods of time and experience pure joy, is the "true" (as I call it) NOS DAC.
Have you listened to NOS DAC's and maybe compared them to other types?
Hi,
In general I also prefer the NOS DAC, there is one factor that is quite often forgotten. Music is a bunch of random signal that is quite often non-repetitive. The problem with the reconstruction filter is that it can insert points that are not in between the two voltage levels, that is what it is suppose to do. That is why it is called a "reconstruction filter". The problem with that is what if the algorithm guesses wrongly, so it could inadvertently insert another signal that wasn't there in the original music. I don't think this area has actually been studied and there is no way of studying it. So while looking at a music signal it might decide that the waveform must have come from a 10KHz signal, when it is not, and creates one. that would have been one nasty "sss" that you would hear..
Oon
Yep, I agree.
If the interpolated point calculated by the OS filter wasn't in the original music then that can only mean the aliasing requirement at the input has been violated. Otherwise the number of possible points it could be is heavily constrained by the input bandwidth restriction. Its not really true to say (as I've heard Mike Moffat say) that the interpolation filter 'guesses' at the missing value. If its a guess then its a highly 'educated' guess.
Of course differently designed filters will give slightly different values for the educated guesses. But then so will analog filters after the DAC, which as E_B has already pointed out, already exist in the downstream system and we can't avoid them.
@Marcel - yes indeed, an HF roll-off definitely 'smooths' the perceived sound and makes things sound more 'distant'.
Well one major difference between analog filters and reconstruction filters (RC) is analog filters do add in a voltage that is higher than the two points. It is not physically possible. For example the 1st point is 5V, the second point is 2V, the in between point must be somewhere between 5V and 2V. Everything is real time.
A reconstruction filter on the other hand, if it guesses that this might have been from two points of a 10 KHz signal but the peak is in between, could insert in a 12V point to create a sine wave.
That is why NOS even with the best analog filters still produces lousy looking 16KHz signal. Because there is not enough points to create the signal properly. The peak is missed. So if you actually played a 16KHz through a NOS DAC, it will appear as a 16KHz signal amplitude modulated at 8KHz (22.1-16KHz). No amount of Analog can save that.
On my personal level, I prefer the mids from a NOS dac, but the highs suck. The OS DAC, the highs are nice but the mids suck...
I have read some articles that says a reconstruction filter with a more gradual slope can be the best compromise between the two...
Oon
Turns out what you're saying isn't borne out in practice. I have a filter which is a 5th order LC (quasi elliptic, much steeper than I've seen elsewhere on another DAC) which does a pretty good job on a 16kHz sinewave - if you (or anyone else for that matter) want to see how it looks I'll take a pic or figure out how the screensave function works on my 'scope and post it up. Analog filters can indeed do the job as well as digital (oversampling) ones - in this case the image freq of 16kHz (which is 44k1-16k0 = 28.1kHz) gets about -30dB suppression.
I agree - but only on NOS without the steep anti-imaging filter. With the filter, the highs are nicely sweet.
I've read that too - most recently in the paper by Bob Stuart. I'm staying in the camp of 'steep is best' for now though.
I don't think that is actually physically possible. Look at it from a physical sense. I actually have no clue about the filter you spoke of. But it would be really interesting to see the waveform before the filter and after the filter.
Let's say for the first point of sampling a 16KHz sinewave is at 0 at the beginning of the waveform. There are a total of 3 samples from the a single 16KHz cycle. The second point would be just before the signal transitions from positive to negative (would be a very low amplitude signal and the third point would be just before negative swinging to positive). The sampling has missed out all the high peaks of the signal. Therefore the NOS DAC would only produce the a low level signal. It would look just like a 22.05KHz being modulated by a 8KHz signal. At one of the later waveforms the signal would only actually capture the peak. Pretty much like an radio AM signal. I am not sure how an analog filter could add in all the missing peaks the signal that the NOS DAC sends out...
Sorry Typo there. Should have read
Well one major difference between analog filters and reconstruction filters (RC) is analog filters can't add in a voltage that is higher than the two points.
Here's a quick snap from my scope of the filtered DAC playing out a full-scale 16kHz sinewave - there's still some modulation effect as you can see, but the signal level isn't reduced according to your hypothesis. In fact with NOS the filter is required to provide some gain as the ZOH incurs about 3.2dB attenuation at 20kHz - this filter incorporates that.
I can't do a before-after filter comparison as with this DAC the filter's prior to I/V meaning there's only current possible to 'scope, no voltage and I don't have a sensitive wideband current probe to employ. I can jury rig another NOS DAC to give a comparison shot though. Will look into that.
I can't do a before-after filter comparison as with this DAC the filter's prior to I/V meaning there's only current possible to 'scope, no voltage and I don't have a sensitive wideband current probe to employ. I can jury rig another NOS DAC to give a comparison shot though. Will look into that.
Attachments
Last edited:
No. No guessing takes place. The reconstruction filter does not know (cannot know) and does not care what the original signal (taken as a whole) was. It simply calculates the next output value from what it has already seen. This area is well-studied and can be found in any good textbook on digital signal processing.oon_the_kid said:
Yes.abraxalito said:
No. You may be thinking of a low order low pass filter, which is essentially just a smoother. Higher order filters (whether digital or analogue) can do what you believe to be impossible.oon_the_kid said:
Unfiltered or poorly filtered 16kHz from a NOS DAC will look like 16kHz plus a smaller amount of 28.1kHz, which may look rather like 22.05kHz double-sideband modulated at 6.05kHz. The reconstruction filter removes the 28.1kHz image to leave 16kHz. A good analogue filter can do this.
That much is clear. You need to do some more reading and some more thinking, then you can be sure. It took me quite a long time to get my head round digital audio; I kept thinking of snags (just like you are doing) but as I gained understanding I realised that there were answers to all of the apparent snags so digital audio actually works as it is supposed to work.
As a first step to enlightenment, stop thinking of a reconstruction filter as a smoother or interpolator. Consider instead an ideal brick-wall filter which can remove the 28.1kHz and leave the 16kHz. Real filters are not that good, but they are not a simple smoother.
Thinking about what a reconstruction filter does in the time-domain (as an oscilloscope would show) isn't particularly helpful to understanding the nature of an NOS or unfiltered signal. Rather, I suggest thinking about it in the frequency-domain (as an spectrum analyzer would show). It's important to recognize that an unfiltered NOS DAC signal FULLY contains the desired signal band information. This becomes clearly apparent when the unfiltered signal is viewed in the frequency-domain. The desired signal doesn't need to be guessed, or created, or even reconstructed, really. Instead, the issue with unfiltered DAC signals is that they also contain repeating copies of the desired signal band information, which are spectrally shifted up in frequency and centered at interger multiples of the sample rate.
The job of the reconstruction filter, whether it's digital or analog, is to remove the repeating image bands leaving only the desired signal band. Viewed this way, you can see that the reconstruction filter doesn't need to guess values in order to reconstruct the original analog signal, it just needs to strip away all of the undesired image bands. The image bands are what give the analog output signal an discrete looking appearance when viewed on an o-scope, even though the DAC's signal is entirely analog even without any reconstruction filtering being applied at all.
Removing the image bands via sharp filtering and the underlying smoothly appearing signal is revealed. Keep in mind that it was completely in there all along, sort of hiding if only viewed in the time-domain. I suggest thinking of the desired signal band more as being revealed by the reconstruction filter, rather than as being created by it.
I think their point is that with fairly high sample rates (96 kHz) and music of which the spectrum already rolls off naturally, you can get away with a relatively smooth filter and still keep the aliasing products well below the noise floor. Of course that means that the choice of anti-aliasing filter becomes dependent on the characteristics of the music; if you would want to record a church organ, you would need to switch to a steeper filter when the bats wake up.
I think the point here is that errors may accumulate over the entire signal chain. Many people would not object against a piece of equipment that rolls off by 1 dB at 20 kHz, but cascade many of them (microphone, microphone preamplifier, ADC, DAC, preamplifier, power amplifier, loudspeaker) and it becomes a different story.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.