[514077]

Your explanation makes for more confusion than it helps. «Taps» is just another term for «coefficients». The number of coefficients stands for the complexity of a filter.

Just as an aside, in reference to other companies mentioning taps, I recall Jason from Schiit mentioning 18,000 taps when refering to the as yet unreleased Yggy.

 

[Mediahound]

Your explanation makes for more confusion than it helps. «Taps» is just another term for «coefficients». The number of coefficients stands for the complexity of a filter.

Not exactly. See this video for Chord's explanation:

 

[rayl]

Though Yggdrasil is an R2R so I suspect the filter in that context is probably for smoothing the high freq noise from the R2R output - one of the drawbacks of R2R as Rob noted. Don’t know for sure. (Confession- I am actually in the Yggy now salivating over DAVE camp after rob’s talk. )

And sorry for my confusing response on taps. What I really wanted to highlight is that it corresponds to the delay... I.e. 100k taps required 100k samples so the output is delayed by 100k samples... though they may (and almost always are) upsamples... bit foggy on what the practical limits on upsampling are but that may be too granular a question to ask here.

And more samples is more complexity as noted. A “perfect” reconstruction requires infinite delay (Or at least as long as the song)... that’s Whitaker Shannon in the slides...

 

[DjBobby]

Your explanation makes for more confusion than it helps. «Taps» is just another term for «coefficients». The number of coefficients stands for the complexity of a filter.

So taps are same as coefficients mentioned in other dac chip specs?
Still not getting it quite clear. Sabre mentions in their es9028pro dac chip specs, "Up to 11 sets of digital filter coefficients to choose". If more taps equals better sound, why shouldn't everybody simply go for more taps aka coefficients instead of offering "11 coefficients to choose"?

 

[Rob Watts]

Since I couldn't find mentioning taps anywhere except for Chord products, are there some synonyms for taps which other manufacturers are using in their specs?

No taps are taps. The tap length (or number of taps, or more technically the FIR filter order plus 1) is not a term that can be confused with something else. Other manufacturers either prefer to minimize the number of taps, because of a mistaken view that pre-ringing is somehow bad, or take the approach of minimum phase, with only post ringing using asymmetric FIR filters (or IIR filters). Both of these approaches will not reconstruct the timing of transients accurately.

The cool thing about listening to the Blu 2 is that immediately after hearing it, you know that it sounds profoundly better - there is no question that it is fundamentally better, as the difference is just too vast. And of course, this is completely backed up by sampling theory. When more people get to hear it, it will become self evident that increasing tap length is the correct thing to do. When it eventually becomes an accepted fact that this is the way forward, then it will be interesting to see how the other manufacturers respond. No doubt we will hear tales of we always knew tap length was important, but didn't advertise our huge lengths. Perhaps we will get pseudo tap length numbers and other marketing nonsense, trying to confuse the situation. But the number of taps are the number of taps...

Coefficients can equal the tap length, but for a ideal sinc function the coefficients are symmetric, so for an even tap length, the number of coefficients stored is exactly half the tap length. Then you have half band filters, where almost half the coefficients are zero...

 

[JaZZ]

Not exactly. See this video for Chord's explanation:

What are you referring to? I can' find any discrepancy to my post – «Taps» aren't even mentioned.

So taps are same as coefficients mentioned in other dac chip specs?
Still not getting it quite clear. Sabre mentiones in their es9028pro dac chip specs, "Up to 11 sets of digital filter coefficients to choose". If more taps equals better sound, why shouldn't everybody simply go for more taps aka coefficients instead of offering "11 coefficients to choose"?

I can't answer your question. Coefficients may not automatically exactly equal Taps, but Taps definitely equal coefficients*. In the case at hand «coefficients» perhaps should mean «characteristics».

why not just use a battery powered laptop with ssd drive. rob has clearly stated this in conjunction with dave's galv.isol. usb offers total isolation better than optical. the money saved could go towards 4K abyss headphones where a massive night/day improvement is instantly noticed?

A battery operated Windows tablet may be what I'll end up with. IIRC the advantage is in the absence of a ground connection via mains cables prone to noise induction.

*I've seen Rob's relativation of my tap=coefficient statement, but it seems that in a broader sense they are used as the same (as found in some Wikipedia articles).

 

[Rob Watts]

Though Yggdrasil is an R2R so I suspect the filter in that context is probably for smoothing the high freq noise from the R2R output - one of the drawbacks of R2R as Rob noted. Don’t know for sure. (Confession- I am actually in the Yggy now salivating over DAVE camp after rob’s talk. )

And sorry for my confusing response on taps. What I really wanted to highlight is that it corresponds to the delay... I.e. 100k taps required 100k samples so the output is delayed by 100k samples... though they may (and almost always are) upsamples... bit foggy on what the practical limits on upsampling are but that may be too granular a question to ask here.

And more samples is more complexity as noted. A “perfect” reconstruction requires infinite delay (Or at least as long as the song)... that’s Whitaker Shannon in the slides...

The delay depends upon the oversampling ratio too. So if it was oversampling by one (i.e. no oversampling, what comes in comes out) then the delay would be the (input sample time)*(number of taps)/(2 * OS ratio).

So for a 16 times oversampled it is:
(input sample time) * (No of taps) / 32

So for the 1M tap Blu with 48kHz it is:

20.833uS * (1,015,808)/32 = 0.661 S

The filter delay is 0.661 seconds, but the filter processes on 1.323 seconds of music.

 

[musickid]

from a non technical background i always understood that the tap number was directly proportional to the ability of the interpolation filter to reconstruct the original analogue signal from the given digital binary 'bits' so the higher the tap number therefore by the very laws of physics the increased accuracy of the reconstructed analogue waveform correlates directly again with relation to the interpolation filter. i can almost picture the nanoscale world whereby all this occurs. has a digital amp been produced by chord yet i ask because i saw an integrated chord digital amp but i also read this was some way off yet so i'm confused. finally will the davina mscaler function exactly like blu2 and will it be on the market in the next two years priced lower than blu2?

 

[Rob Watts]

Actually, Blu 2 used the code that was originally intended for Davina... That code won't change for Davina.

 

[musickid]

it will be fascinating to see how all these developments map out over the next year to two. cheers and thanks Rob.

 

[JaZZ]

...What I really wanted to highlight is that it corresponds to the delay... I.e. 100k taps required 100k samples so the output is delayed by 100k samples... though they may (and almost always are) upsamples... bit foggy on what the practical limits on upsampling are but that may be too granular a question to ask here.

And more samples is more complexity as noted. A “perfect” reconstruction requires infinite delay (Or at least as long as the song)... that’s Whitaker Shannon in the slides...

I've fallen into the same trap some time ago, thinking the number of taps was the number of samples used for calculating the reconstruction of the original signal. But taps are just an indication of the filter complexity. In fact the «reconstruction» consists of a sophisticated interpolation between the samples with the goal of reconstructing the low-pass filtered signal that has reached the A/D converter, not the original analogue signal, which may have contained some ultrasonics leading to a relatively jagged signal shape. That's why low-pass filtering prior to A/D conversion is essential – for suppressing aliasing. Unfortunately – since Davina isn't studio standard yet – this anti-aiasing filtering already implies some transient corruption.

Conventional «reconstruction» (low-pass) filters can't avoid pre- and post-ringing with effects down into the audio band (the infamous timing issue), whereas a «perfect» reconstruction consists of a filter algorithm without pre- and post-ringing comprising audible frequencies, but infinite ringing at the filter frequency instead. Chord's DACs, particularly the BluDave, are successfull approaches to the theoretical ideal.

 

[rayl]

In the context we are using here, it does correspond (linearly) to the # of samples.... Half from the "future" (i.e. delay needed) and half from the past as the filter is centered in the time domain..... Though the samples being counted may be the result of oversampling (I think I said upsampling, but I meant the same thing), so it may not be actual PCM samples per se but a linear multiple, e.g. as in Rob's 16X oversampling example, 1 actual sample may count as 16 oversamples for the purposes of corresponding to the tap count....

Oversampling introduces guard bands in the frequency domain, effectively reducing the impact of a less than ideal filter (less than a perfect rectangle in frequency domain). A picture of this can be found on page 11 of the following I found via Google http://www.ece.iastate.edu/~namrata/EE424/l1.pdf -- but of course, there are limits to oversampling (an area I am much less familiar with).

I've been quite inspired by the 2 talks to brush up a bit on signal reconstruction. They were great talks. Finally something to cut through the marketing euphemisms and noise. The comment about how DSD misses certain HF transients until the errors accumulate enough was also enlightening -- I've never been comfortable with DSD in principle, though some would argue it simplifies the analog side of reconstruction and it's always easier to implement better math than better circuits. Hearing a DAVE was the first task upon returning from RMAF and I am impressed. Now, just need to upgrade to one (or maybe a successor).

 

[musickid]

very simplistically put i also remember that the tap count say of 49K for hugo 2 only indicates the number of taps in the first stage of the "filter"??? the real tap count is not published commercially and may be far higher. correct me if i'm wrong as i can only understand this from a qualitative perspective not being technically trained. it is satisfying to possess a basic understanding of the technology you have acquired.

 

[Rob Watts]

I've fallen into the same trap some time ago, thinking the number of taps was the number of samples used for calculating the reconstruction of the original signal. But taps are just an indication of the filter complexity. In fact the «reconstruction» consists of a sophisticated interpolation between the samples with the goal of reconstructing the low-pass filtered signal that has reached the A/D converter, not the original analogue signal, which may have contained some ultrasonics leading to a relatively jagged signal shape. That's why low-pass filtering prior to A/D conversion is essential – for suppressing aliasing. Unfortunately – since Davina isn't studio standard yet – this anti-aiasing filtering already implies some transient corruption.

Conventional «reconstruction» (low-pass) filters can't avoid pre- and post-ringing with effects down into the audio band (the infamous timing issue), whereas a «perfect» reconstruction consists of a filter algorithm without pre- and post-ringing comprising audible frequencies, but infinite ringing at the filter frequency instead. Chord's DACs, particularly the BluDave, are successfull approaches to the theoretical ideal.

The ideal reconstruction filter (sinc function) will perfectly reconstruct the bandwidth limited analogue signal that was in the ADC, before it was sampled.

So the real issue is whether bandwidth limiting has an impact on sound quality. For sure, inadequate bandwidth limiting has a huge effect on sound quality - I know this from designing and listening to decimating DSD filters - and the problem is down to aliasing. I have designed the first decimation filter for Davina, and this has 260 dB rejection, so this will ensure aliasing is in practice well below -300dB. This uses symmetric filter, so will pre-ring. I am working on a IIR type filter that also has similar aliasing but of course no pre-ringing.

These filters have been designed as non decimating too, so we can listen to the effects of bandwidth limiting directly. I am hopeful that Davina's out of band output will be so low, that the benefits of bandwidth limiting won't outweigh any potential downside to bandwidth limiting. But anyway, we shall soon be able to hear the effects.

 

[DjBobby]

No taps are taps. The tap length (or number of taps, or more technically the FIR filter order plus 1) is not a term that can be confused with something else. Other manufacturers either prefer to minimize the number of taps, because of a mistaken view that pre-ringing is somehow bad, or take the approach of minimum phase, with only post ringing using asymmetric FIR filters (or IIR filters). Both of these approaches will not reconstruct the timing of transients accurately.

Thanks a lot for the taps clarification.
I watched with a pleasure the entire RMAF 2017 video, trying as a layman to understand (more or less) the principles. Only one thing left me clueless - the section with the "pre-ringing is not bad". Especially the statement "the more pre-ringing, the more accurate the reconstruction is." From my personal experience, whenever i had a choice of filters, the sharp brickwall filter was the least desirable one, being the most fatiguing. Also the Ayre white papers talks about pre-ringing as time-smear and slow roll-off filter improving transient response. I would greatly appreciate if you could drop a few more words about this part of your presentation.