Digital that sounds like analog

[LL21]

The common thing with all the people doing NOS dac is that the digital filters is the main problem, it seems. Removing the filter function results in better SQ hearing wise.

Going back to Peter, he talks about frequency and time domain and feels that not a lot of thoughts/effort has gone into the timing side of things, which ties in with Opus's thing with CM distortion. Lately I have also find that phase and polarity does affect SQ to a large degree but it is a subject that most people dismiss or ignore.

Your last statement is intriguing to a non-techie (me) in that the Zanden NOS DAC made a big point about no digital filter, but also about its emphasis on phase linearity which Yamada San felt was quite important (he also has a polarity switch as one of 3 main knobs on the DAC). Are these 2 the same as far as you know as to what you were describing in your post? (dont want to confuse terminology). If yes, can you share more of your observations about what makes these 2 elements (Phase and Polarity) important to you in digital?

 

[opus111]

He certainly is succinct, but he's 100% wrong in the first sentence, so I have given up. He says:

No analog filter can replace the digital filter.

Sounds to me like he could be offering here a plug for his own product. My Ozone DAC has no digital filter to speak of, and it does have an analog filter.

 

[bbb]

Phase relates to timing and polarity, directionality. Most digital filters affect phase. NOS, due to the filterless feature is perfect in the time domain but suffers in the frequency domain (3db droop and aliasing/imaging effect). Polarity affects the timing indirectly due to the reversal of compression and rarefaction. I can hear this effect but a lot of people don't so YMMV. Peter's solution is to do the filtering in the computers so the dac is techincally NOS but since his filter needs 768khz to work effectively, the problems of upsampling crops up, but again, his upsampling is handled by the computers via XXHigh End. So that could be the key. I am still trying to digest that part so maybe Opus can chime in first.

 

[opus111]

As far as I can work out (and I can't understand Peter's prose) he's using time domain interpolation (rather than frequency domain). Which means when he calculates the 15 new samples to place in between the original ones - now spaced 16 samples apart with the 768kHz sample rate - he's going to generate frequency domain distortion. That in addition to making his PCM1704s work 16X as hard - if you check the datasheet you'll notice the measured figures degrade at higher SRs. So the only benefit I can see here is bigger numbers - a marketing benefit for sure as probably the majority of customers still think bigger numbers mean better.

 

[bbb]

He certainly is succinct, but he's 100% wrong in the first sentence, so I have given up. He says:

No analog filter can replace the digital filter.

Sounds to me like he could be offering here a plug for his own product. My Ozone DAC has no digital filter to speak of, and it does have an analog filter.

So, leaving aside his bold claim, are you saying that his reasoning on why an analog filter will not be as effective as his digital filter is not correct?

<edit> wrt to post 385, leaving aside the negative effect his upsampling and filtering method does to the 1704s, will you still say the above is correct?

 

[opus111]

I didn't get to the reasoning - gave up when I found the claim to be nonsense. When something's obviously nonsense I'm not interested in justification for it An analog filter is required for all DACs to attenuate image frequencies - a digital filter can't ever replace an analog filter, but it can make it easier to design in theory. This at the expense of running the DAC chip faster and hence getting worse sound due to increased glitching.

 

[bbb]

Ok let me try again. S-D dac is no good because they upsample and filter digitally. It is easier to construct gentler (better) filter at higher frequency away from music band. NOS is good because being filterless creates no problem in the time domain. NOS is bad because of problems in the frequency domain (droop and aliasing/imaging(filterless)). So transpose to the phasure, filter at 768khz is best due to no or least problem in time domain (peter's assertion not mine). Due to filter requirement at 768khz, no choice but to upsample but do both upsampling and filter at computer so to minimise problems at dac side. Upsampling creates problem for dac but probably less than if done on dac. Many people still prefer NOS despite problems in frequency domain so maybe our hearing is more sensitive to time domain problems. Overall, the upside in maintaining time domain integrity and proper treatment of aliasing/imaging outweighs the downside in the frequency domain created by running the dac at 768khz.

edit <NOS is good because it has less or no common mode distortion. common mode distortion is more of a time domain problem since it seems to affect timing qualities in music, ie. dynamics and musicality and thus affect S-D dac more. common mode distortion is a time domain problem because it rarely/doesn't show up in frequency domain measurements>

Please feel free to tell me where I have gone wrong. Just trying to learn.

 

[opus111]

Ok let me try again. S-D dac is no good because they upsample and filter digitally.

No, what's wrong with S-D isn't the upsampling or the digital filtering - both are benign. What's going wrong in S-D is inadequate dither being used when truncating the input data. Truncation is a grossly nonlinear process. Feedback isn't the way to linearize a dead band in a transfer function.

It is easier to construct gentler (better) filter at higher frequency away from music band. NOS is good because being filterless creates no problem in the time domain.

I think the phase performance of NOS is purely incidental to its great sound.

NOS is bad because of problems in the frequency domain (droop and aliasing/imaging(filterless)).

Those are the two weaknesses of NOS that designers/manufacturers seem to ignore, in general. Perhaps because they're subscribing to the notion that its those factors which give NOS its feature sound.

So transpose to the phasure, filter at 768khz is best due to no or least problem in time domain (peter's assertion not mine).

Remains merely an assertion as far as I can see. No evidence that time domain performance is holding back digital that I've seen.

Due to filter requirement at 768khz, no choice but to upsample but do both upsampling and filter at computer so to minimise problems at dac side.

Shaky premise, therefore unsound conclusions.

 

[Orb]

He certainly is succinct, but he's 100% wrong in the first sentence, so I have given up. He says:

No analog filter can replace the digital filter.

Sounds to me like he could be offering here a plug for his own product. My Ozone DAC has no digital filter to speak of, and it does have an analog filter.

Just to add and agree to your comment, I think T+A use analogue reconstruction filter and the extensive Hifi News measurements show it is an excellent engineered product with comparable SOTA results (apart from niggle with USB), I assume Peter's context is reconstruction filter?
Cheers
Orb

 

[bbb]

No, what's wrong with S-D isn't the upsampling or the digital filtering - both are benign. What's going wrong in S-D is inadequate dither being used when truncating the input data. Truncation is a grossly nonlinear process. Feedback isn't the way to linearize a dead band in a transfer function.

Ok. So how does this relates to common mode noise which is what seems to plagued S-D dacs most.

 

[opus111]

Well if you look up dither theory you find that the result of inadequate dither is noise modulation. Noise modulation occurs if the dither isn't TPDF - which seems to be the case with S-D DACs as its practically impossible to control when there's so much feedback. Noise modulation isn't the same as common-mode noise though - that's a system phenomenon.

 

[Orb]

No, what's wrong with S-D isn't the upsampling or the digital filtering - both are benign. What's going wrong in S-D is inadequate dither being used when truncating the input data. Truncation is a grossly nonlinear process. Feedback isn't the way to linearize a dead band in a transfer function.



I think the phase performance of NOS is purely incidental to its great sound.



Those are the two weaknesses of NOS that designers/manufacturers seem to ignore, in general. Perhaps because they're subscribing to the notion that its those factors which give NOS its feature sound.



Remains merely an assertion as far as I can see. No evidence that time domain performance is holding back digital that I've seen.



Shaky premise, therefore unsound conclusions.

Opus,
when you mention S-D and digital filtering being benign.
Hmm so what about pre-post ringing and ripples/filter algorithms better in either but not both time and frequency domain, or are you talking about digital filters in a very specific context?
I appreciate the aspect I am raising is not just restricted to S-D.

Thanks
Orb

 

[bbb]

Opus,
when you mention S-D and digital filtering being benign.
Hmm so what about pre-post ringing and ripples/filter algorithms better in either but not both time and frequency domain, or are you talking about digital filters in a very specific context?
I appreciate the aspect I am raising is not just restricted to S-D.

Thanks
Orb

pre-post ringing will actually affect time domain, no?

btw, the paper that myles' post, http://www.whatsbestforum.com/showthread.php?9582-Math-Unraveled, supports my assertion that human hearing is more sensitive to timing rather than frequency difference. there is a section where they recorded a segment of dialogue from the movie casablanca where they compare one with the phase information is destroyed and one where phase info (timing) is retained but white noise is introduced (frequency) which i think is relevant to what we are discussing wrt to NOS, S-D, digital filter etc.

 

[DonH50]

Both digital and analog filters exhibit ringing, although no pre-ringing in analog (causality). Ringing leads to ripples in time and frequency.

Both types of filters affect phase, it is impossible to not (always some delay in the real world). There are linear phase filters that can be implemented using analog or digital circuits. Linear phase over frequency means constant group delay; that is, signals at every frequency are delayed equally, so a pulse coming out is identical to the one going in except for the time delay.

I am pretty sure it has been known we are more sensitive to time than frequency for some time; the paper Myles cited demonstrates cases where the uncertainty principle is violated by our hearing. That is, it says we can discern time differences better than the theory says. There are debates about that, of course, but the arguments are well beyond this thread and probably beyond me -- at least, I would have to study it. I have dealt with the effect, but resolved a similar issue with processing. Or rather our signal-processing gurus did, I was just the hairy-knuckled engineer building the stuff.

 

[opus111]

Hmm so what about pre-post ringing and ripples/filter algorithms better in either but not both time and frequency domain, or are you talking about digital filters in a very specific context?

I was talking about filtering as a process being a benign one, but its possible to screw up the sound with certain implementations yes. I don't think post-ringing is much of an issue if the ringing frequency is high enough. Ditto pre-ringing though I'm more uneasy about that as its unnatural. Filters can have other issues - pre-echo is one I'd definitely want to avoid, which plagues some FIR equiripple implementations. Aliasing is another one best steered clear of, so this means avoid half-band filters (meaning almost all the commercially available digital filter chips which use this to save on silicon real-estate).

 

[bbb]

Thanks Don for the explanation. So perhaps with our more than previously thought highly sensitive time domain hearing, any deterioration, however slight, in phase information wrt to filters could lead to the the detectable lower SQ and that this could possibly be corrected via processing either through dsp or computers provided these are transparent enough not to impose its own distortions? But wouldn't these dsp or computers be considered as another filter?

 

[DonH50]

Yes, and yes. DSP can create as many ills as it solves if not applied correctly. Just like fire and so many other useful things.

Semi-random thoughts:

DSP is great for implementing very complex filters that are stable and insensitive to component tolerance, temperature, power supply variation, etc. For many applications they provide a giant leap forward in signal processing.

One of the biggest issue with early DACs (and ADCs) that were all NOS back then, was the impact of the steep anti-imaging (anti-aliasing for the ADC) filter required to suppress out-of-band signals. Those filters were of very high order and caused time- and frequency-domain artifacts extending well down into the audio band. opus111 is working on a better filter design, I believe, that is analog yet linear phase. The principles have been around for ages, but such filters are harder to design and generally cost more than the simple filters manufacurers used.

Oversampling DACs get around the filter problem by moving the image/alias frequencies well above the audio band, so simpler (cheaper) filters can be used without corrupting the audio passband. Note any DAC can be oversampled, but delta-sigma (DS, so-called "1-bit" though most all today actually use a mutibit loop) modulators took the idea to the extreme of very few bits (can be just one) sampled at a very high rate. Filters then convert that low-resolution but very fast data stream into much slower, high-resolution signals.

I am not sure the audibility of some of the artifacts seen in today's filters, but a number of reviewers claim to prefer one over another, and the current trend seems to be a combination of types to provide the best time and frequency response.

 

[bbb]

One of the biggest issue with early DACs (and ADCs) that were all NOS back then, was the impact of the steep anti-imaging (anti-aliasing for the ADC) filter required to suppress out-of-band signals. Those filters were of very high order and caused time- and frequency-domain artifacts extending well down into the audio band.

Oversampling DACs get around the filter problem by moving the image/alias frequencies well above the audio band, so simpler (cheaper) filters can be used without corrupting the audio passband. Note any DAC can be oversampled, but delta-sigma (DS, so-called "1-bit" though most all today actually use a mutibit loop) modulators took the idea to the extreme of very few bits (can be just one) sampled at a very high rate. Filters then convert that low-resolution but very fast data stream into much slower, high-resolution signals.

Ok. So a filter with a corner frequency of 44.1khz has no chance of not having time and frequency domain artifacts? Further, a high sampling rate and gentler filter is the best we can do atm? Will the distortion arising from operating multibit dac at higher sampling rate affect the time domain?

Phasure claims that their filter with a corner frequency of 176khz combined with 768khz oversampling creates no ringing and results in only 0.0018 thd+n. Could this be achieved or is it just marketing talk?

 

[opus111]

Even if the DAC's anti-imaging filter has no ringing, ringing has already been baked-in because digital requires band-limiting prior to the ADC. This anti-aliasing filter must have ringing, and its non-negotiable - so ringing still exists irrespective of the interpolation algorithm. Seems daft to me to focus on the total elimination of ringing at the DAC. The THD+N claim is percent or a fraction?

 

[DonH50]

Ok. So a filter with a corner frequency of 44.1khz has no chance of not having time and frequency domain artifacts? Further, a high sampling rate and gentler filter is the best we can do atm? Will the distortion arising from operating multibit dac at higher sampling rate affect the time domain?

Phasure claims that their filter with a corner frequency of 176khz combined with 768khz oversampling creates no ringing and results in only 0.0018 thd+n. Could this be achieved or is it just marketing talk?

Any filter is going to impact time and phase, the only questions are how much and (in this case) is it audible? Some filters work better than others, presumably designers choose what sounds best in their systems, using a combination of digital and analog filtering.

I do not know what "atm" means in this context (not automatic teller machine??? ).

A higher sampling rate menas lower-order filters can be used, but as opus111 has noted introduces other issues, including higher noise, and potentially larger glitches and settling issues related to the higher rate. Time and frequency domain are two sides of the same coin; one will affect the other.

I think you may be overstating how much distortion we are discussing here -- for virtually any modern DAC it is inaudible to all but a select few. All in life and DACs is a compromise.

As for Phasure's claims I have no way of verifying, but it is certainly possible to design a filter that does not ring, they are used all the time. Oversampled or not (you probably give up out of band rejection). Note ringing in (any) filter is usually linear and thus not increase THD. The claims could be easily verified by someone with the equipment and time to test , e.g. JA at Stereophile.