Why do NOS dacs sound different to oversampling designs?

[Opus111]

With oversampling the high slope filter is still there - normally that's the first interpolate-by-2 filter in the chain which needs a high slope to reject anything over 22.05kHz in the original sample stream. Its just that its now a digital filter so its not subject to the vagaries of component tolerances and drift. Its normally made linear phase too, something very difficult to do with analog.

 

[Ken Newton]

...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.

I would be interested in viewing that presentation if you can locate it, Richard. Thanks.

 

[Ken Newton]

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

No worries.

 

[DonH56]

In a normal system there can't be anything over 22.05 kHz going to the DAC -- it was bandlimited by the ADC and processing before the DAC.

Note that oversampled converters are not always delta-sigma designs (my comment, not something Opus111 implied).

 

[Opus111]

I would be interested in viewing that presentation if you can locate it, Richard. Thanks.

Its here, a fairly long .pdf, the relevant pages are 49 through 63 : http://www.hypex.nl/docs/papers/AES123BP.pdf

 

[Werner]

I do not see how not suppressing images can affect timing

The images not being suppressed is the same as saying that the original signal is not being reconstructed. And when this reconstruction fails, or is absent, how can you expect timing to be correct?

But this is moot.

When dealing with music there is always a listener at the end, and this listener has two reconstruction filters in the ears. So, after all is said and done, even an unfiltered NOS signal hits a steep lowpass filter at < 20kHz, and this filter performs the reconstruction (and spectral distortion) of the original signal, in its pass band. Including timing.

 

[Purité Audio]

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.

I think I almost understood the above, Werner ,nice to see you here by the way, are you saying that no reconstruction filter is needed?
In simple terms is there an 'ideal' method to transcribe the original analogue signal ADC/DAC back into analogue?
Keith

 

[Werner]

In simple terms is there an 'ideal' method to transcribe the original analogue signal ADC/DAC back into analogue?

Forget about the ADC side: there are no rules there. Assume that the signal, by magic, gets bandlimited and sampled.

Then ask: " is there an 'ideal' method to transcribe the original sampled signal, via a DAC, back into analogue?"

The answer is YES. How to do it is literally described in the (proof of the) sampling theorem, and this is
implemented in our oversampling digital filters with varying degrees of accuracy (or deliberate inaccuracy).

It is 2016. Strange that this still has to be spelled out every once in a while...

are you saying that no reconstruction filter is needed?

The ideal reconstruction filter is the Sinc(x) function, i.e. a zero-ripple linear phase brickwall at the original Fs/2.

But the output of that filter is listened to by an observer through human ears. These ears are mechanical lowpass
filters with, give or take, infinite attenuation even before Fs/2 is reached. In other words, the ear-filter cuts out even more aggressively than the preceding reconstruction filter. This makes the reconstruction filter irrelevant, under the assumptions of this discussion. More formally: the cascade of a Sinc filter at 22kHz with an ear at 20kHz is equivalent to an ear at 20kHz.

The assumptions are:
-zero-width samples during replay
-perfectly linear amplifiers and speakers.

Both of which do not hold in the real world, and as such these introduce minor deviations from the ideal case.

 

[Purité Audio]

Thanks W,
'The ideal reconstruction filter is the Sinc(x) function, i.e. a zero-ripple linear phase brickwall at the original Fs/2.'

One last question, is the ideal reconstruction filter then the inverse of the original filter used in the ADC?
thanks for your patience.
Keith.

 

[Werner]

One last question, is the ideal reconstruction filter then the inverse of the original filter used in the ADC?

No.

As I said before: there are no rules for the original sampling and the anti-aliasing filter. The sampling theorem assumes that the samples appear by magic and proceeds from this.


But something to ponder ...

Assume we were defining a digital audio standard for 22kHz sampling, i.e. with a paltry bandwidth of 11kHz and a cut-off frequency
in the audible range of most not-too-old humans.

We know that a Sinc at 11kHz is perfect for reconstruction/anti-imaging (DAC). So we'll use that.

But if we were to apply a Sinc at 11kHz for anti-aliasing (ADC), it would likely sound pretty dire: the pre-ringing of that filter would be readily audible.

 

[DonH56]

The images not being suppressed is the same as saying that the original signal is not being reconstructed. And when this reconstruction fails, or is absent, how can you expect timing to be correct?

But this is moot.

When dealing with music there is always a listener at the end, and this listener has two reconstruction filters in the ears. So, after all is said and done, even an unfiltered NOS signal hits a steep lowpass filter at < 20kHz, and this filter performs the reconstruction (and spectral distortion) of the original signal, in its pass band. Including timing.

Hmmm... Of course not suppressing the images contributes signals outside the bandwidth of the original signal, but do those images corrupt the timing of the baseband signal? I am not sure I see that but admittedly have not spent time thinking about it. Those images are sometimes useful in the RF world (saves a mixer) but they are nothing but trouble in a baseband (audio or otherwise) systems.

On the DAC output sinc (sin(x)/x) filter; a lot of DAC's I have seen/measured use a peaking filter (digital before or analog after) to prevent the normal -3.54 dB sinc rolloff at Nyquist when sampling at 44.1 kS/s. That adds its own set of issues, natch.

Also natch, I agree with the rest of your posts, not that you need my opinion!

 

[Opus111]

Hmmm... Of course not suppressing the images contributes signals outside the bandwidth of the original signal, but do those images corrupt the timing of the baseband signal?

It all depends on the meaning of 'baseband signal' here. If you only are considering 0 - 20kHz then the images don't reside in that bandwidth and hence don't change any signal in that bandwidth. Bruno's pictures which I linked to are showing what happens when there's zero filtering applied - the ZOH function having infinite bandwidth.

 

[Ken Newton]

...Bruno's pictures which I linked to are showing what happens when there's zero filtering applied - the ZOH function having infinite bandwidth.

Since we don't have benefit of the audio which goes with his presentation, I will draw a presumptive conclusions about one of the the points that appears indicated by slides 74 through 76. Bruno uses the phrase Time-Of-Arrival (TOA) when referring to the time resolution of a sampling system. (and, appearing to draw the opposite comclusion that Bob Stuart does in his afore mentioned AES paper. Bruno makes the well known point that half-band FIR filters violate Nyquist. He goes on and appears to suggest that this violation can be corrected simply by including a single filter that fully reaches it's stop-band by the Nyquist frequency SOMEWHERE in the ADC-to-DAC chain. I do not see how this can be true regarding ADC aliasing. As far as I know, once aliased frequencies are folded in to the desire signal band they are there to stay.

However, I do suspect that DAC signal reconstruction SINC filters might benefit from a pass-band narrowed just enough to place the stop-band at the Nyquist frequency. Interestingly, that is a part of the approach Peter Craven suggests in his AES paper on Apodizing filters for audio. Craven, however, goes further to recommend a minimum phase type filter (having only a post-ringing impulse response) to 'correct' for all of the prior linear phase half-band filters.

 

[Werner]

Bruno makes the well known point that half-band FIR filters violate Nyquist.

A salient point is that the ideal filter, Sinc, also is a half-band filter.

The difference with real-world HBs of course is that Sinc is infinitely long, hence infinitely steep, and so has a stop band of width zero: Shannon is violated, over a spectral width of ... zero.

He goes on and appears to suggest that this violation can be corrected simply by including a single filter that fully reaches it's stop-band by the Nyquist frequency SOMEWHERE in the ADC-to-DAC chain. I do not see how this can be true regarding ADC aliasing.

This is entirely in the context of an ADC or of downsampling. Practical HB filters have non-zero attenuation at Fs/2, and so the upper half of their transition band aliases into the baseband. Assume CD rate and that the transition band extends to 24kHz. Then the 22-24kHz information aliases, corrupts, the 20-22kHzpart of the baseband. This section of the baseband no longer has correct time or arrival, simply because it contains the original signal, plus something else. Since it is not correct anymore, it might as well be cut out of the baseband.

Hence the call for a single filter in the signal path that reaches its stopband at, say, 20kHz. The best place for this filter is at mastering, removing any aliased bands from previous production steps, and avoiding any future imaging in the HB filters of the replay DAC. This 20kHz filter is to be steep, because Putzeys' opinion is that ringing is not the problem.

Craven's strategy looks similar, but he prescribes a shallow filter, exactly because he feels the ringing is the culprit. Even when that ringing is at ultrasonic frequencies (which IMO is bonkers: pre-ringing is easily heard when the filter cutoff is in the audible band, but the typical artefact of it vanishes when the cutoff is increased beyond the person's upper limit).

On the other hand, against Putzeys' reasoning goes the notion that if a person cannot hear beyond 20kHz, then surely any aliasing above 20kHz is not detectable.

 

[Ken Newton]

This is entirely in the context of an ADC or of downsampling. Practical HB filters have non-zero attenuation at Fs/2, and so the upper half of their transition band aliases into the baseband. Assume CD rate and that the transition band extends to 24kHz. Then the 22-24kHz information aliases, corrupts, the 20-22kHzpart of the baseband. This section of the baseband no longer has correct time or arrival, simply because it contains the original signal, plus something else. Since it is not correct anymore, it might as well be cut out of the baseband.

Werner, thanks for that clarification. I should have seen it myself, but somehow was overlooking the obvious.

 

[Ken Newton]

...On the other hand, against Putzeys' reasoning goes the notion that if a person cannot hear beyond 20kHz, then surely any aliasing above 20kHz is not detectable.

This touches on the key DAC design issue I'm interested in at the moment. If I correctly understood what Bruno was suggesting in the presentation, he's saying that the unsuppressed (near) ultrasonic alias products are audibly altering the reconstruced signal in the time domain even while those same alias products are themselves inaudible in the frequency domain.

It appears that Bruno and Bob Stuart agree that the signal time domain is important, yet propose completely opposite approaches to accurately construct it. Bruno, via an in-band, SINC impulse repsonse brickwall reconstruction filter intended to remove encoded HF aliases. Stuart, via an wide transition-band, aperiodic impulse response reconstruction filter. Seems logical that they can't both be right, can they?

 

[abm0]

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".

They must've been reorganizing the website, the paper is now available here: http://www.aes.org/e-lib/download.cfm/17501.pdf?ID=17501

 

[solderdude]

My avatar just could not help himself and says .. 'hi skipper !' (Ken Newton's avatar) while enjoying J.J.Cale's new 2019 album !

 

[abm0]

So what I'm getting so far, based on my limited understanding of the processes involved, and on what little I remember of the basics of signal theory, is that Filterless designs should sound better because they reproduce ultrasonic content, which is necessary for the correct presentation of ultra-fast transients. The human perceptual system apparently can detect transient phenomena on the order of 5 us (which requires the presence of frequencies up to 200 kHz in the frequency representation), even though it can't detect continuous sine waves at such high frequencies.

Am I getting this right? Is this the basic (plausible) explanation for Filterless DACs sounding different?

 

[solderdude]

No... the ultrasonic content you get is NOT harmonically related, in fact it is anything but harmonic content.
The reason for audible differences is the 'treble roll-off' which isn't a treble roll-off like any normal filter.
The reason is that a filterless DAC does not comply to the sampling theorem and that's why it sounds different (to some better, to others rolled off, to others a bit 'gritty', to others 'more dynamic')

Here is a similar discussion