Where has the NDX into Hugo thread gone?

Mark I certainly feel if your system needs to have a source with a kick in it with sparkle, and perhaps a   less tightly resolved bass then the Hugo would not be for you.. But in all honesty if you want added punch and bite i do think the NDAC/555PS is hard to beat. It worked for me until my tastes evolved.

Simon

a potentially highly confusing post of yours, but the LP in my experience was certainly outclassed in musical terms by RED book CD from the early 90s if not actually sooner.

ATB from George

Hi Mark -

It's cool that you are working through the concepts and theory of signal reconstruction from discretely sampled data sets. I'd advise a bit of caution and balance between the concerns you are raising for yourself by what you are learning and what you enjoy or prefer when your hear various real-world implementations of this scientific art. One famous joke (well, famous in school at least!) is that in theory there's no difference between theory and practice, but in practice there is! 

You raise legitimate concerns when you ask how the higher frequency interpolated data can be known to more faithfully capture the original signal than the primary sampled data. In a previous post, I pointed out that for random data you can't do better at all. For deterministic data (where you know the functional form of the source signal - a pure sine wave for example, but any known function works too) then you can be perfect - and since you have the function you don't need to refer to the sampled data when "filling in the gaps". In between these two limiting cases, you can still add "good" data points at higher frequency and different algorithms for re-using the primary sample data can be more or less accurate. The only way to test the success of a given algorithm requires comparison of the reconstructed signal with the original...and the work of "knowing" (trusting really) that a given implementation is higher or lower fidelity to the original, and under what circumstances (onset of bass notes, preservation of relative amplitudes for different frequencies, whatever) takes a lot more time and effort than just digitizing a single signal and then reconstructing it once on the fly. There really is rather a lot of signal processing out there - and I'm pretty sure you could build yourself an excel spreadsheet to check out a few basic approaches with a few different interpolation filters...it's kinda fun and hugely instructive. 

A couple of posts, including a recent one from you, have said that zero samples can't add information (or words to that effect) - and that you'd prefer your signal raw with no added sugar. Empirically, and almost purely based on your presence in this forum and your choice to buy expensive hifi gear, I'd say you prefer the clever engineering more than the naive straight digital to analog signal! If you wanted to, you could save a bundle and go back to a first generation CD player and re-experience the good old days!

Naim, and most other sophisticated product manufacturers (including Chord, he added to ensure thread relevance) have decided to produce an output signal with a much higher output clocking frequency than the input signal sampling rate. I think I remember seeing that Naim uses forty times higher output clocking than the standard CD red book value (but I'm vague on this and too lazy to surf and find it since this interface ditches my typed text if I go to a different browser tab). (((EDIT: naim does a bit more than 16X to 768 kHz; the 40 I remembered is the bit depth of their processor, more than 2X original for CDs...sorry for the error, it's late here!))) They do this not (only!) for mathematical joy and rigour, but because it is a successful way to reconstruct a more faithful, and hence better sounding, output signal. Really, it works. Really, a lot of different mathematical approaches work too, some better than others, some more pleasing to some, others more pleasing to others. But generally, closer to the (unsampled portions of the) original sounds better and is more pleasing. 

Most algorithms start by zero filling all the new, higher frequency, "dots" for the output, and keep the original data in their spots. Then any convolution style filter will mathematically add up a weighted sum of a given number of data points and update a zero value to the interpolated value. This is obvious if we just want to double the frequency and use a simple linear interpolation: the "one and a half" point will start out as a zero, and will be updated to be the arithmetic mean of the first ("one") and second ("two") points. No magic, no assumptions other than the guess that the local function that was valid between time one and time two is a straight line. In the unlikely event that this is the correct function, then our interpolated point is exactly what the missing sample would have been if we went back and re-digitized at double the speed. Yay. In the equally unlikely event that we had random noise and so there is no relationship between the points at times one, one-and-a-half and two, then our interpolated value is almost certainly wrong and possibly wildly so. Boo. In general, though, for complex but non-pathological signals, like music, this interpolated point will be closer to the original than either the first or second point taken alone. Fancier algorithms use more points at a time and adjust the weights (we used equal weights in the average above) to capture sharper features. Some approaches re-use interpolated data (either single-sided, using the 1.5 point when calculating the 2.5 point, or double sided using both the 1.5 and 2.5 points to calculate the 1.75 and 2.25 points...sort of) and some don't. Few, if any (but I'm not an expert), start off with non-zero filling - largely because that's a much harder assumption to generate in a meaningful way. So it is unlikely that Chord, or Naim, just "invent" interpolated data...they don't have to because they can "invent" the filter shape for the convolution and the scheme for how they will double and re-double the clock rate to build the interpolated waveform. 

I hope this helps ground your arithmetical thinking in some quasi-visualization. i find it useful to think through things this way. But I find it fun to listen and marvel at the results of the clever people who have done all this for me and put it in a nice box! Please let me know if I've missed the point or gone down a wrong road here. If not, we can walk through a scheme to decide how to use adaptive filters to get faithful transient recovery with higher output clocks than the original sample rate. As a quick example, though, consider the (2-D) image processing technique that leaves only edges (often shown in black and white) as an extreme way of interpolating for triggers. 

Best wishes. Regards alan

You're talking classical of course G - and I mainly agree with you on that

The answer is you can only say with certainty it starts at the next sampling point on the CD. You cannot manufacture greater precision on the start point to say 4 microseconds of accuracy. Hypothetically the 'something ahead' could actually start anywhere from 1 microsecond to 21 microseconds into the 22 microsecond period.

Your argument does not even hold for the envelope of a wave let alone the start / stop.

As for your comment above that "I don't know that it is perfect at the infamous 4 us [microsecond] level"......well, Alan...that is the whole point my friend. Glad we agree on this

ATB, MM

Guys, all that increasing the 'taps' does in an FIR reconstruction sinc filter is to improve the responsiveness of the low pass anti aliasing filter with less frequency distortion and artefacts. 

Simon

Mark

I fail to understand why you’re getting in such a tizzy regarding the 4 microsecond remark.

Jan has answered your question regarding the claim of perfect reconstruction. IE according to Rob Watts, to attain perfect reconstruction an infinite number of taps is required.

In fact, Rob Watts has stated that he believes increasing the number of taps beyond the 26k used in the Hugo, will bring about audible improvements. This obviously indicates he does not think the Hugo can reconstruct within 4 microsecond.

You’re arguing against a claim that hasn’t been made.

Guys, all that increasing the 'taps' does in an FIR reconstruction sinc filter is to improve the responsiveness of the low pass anti aliasing filter with less frequency distortion and artefacts. 

If you had an infinitely long piece of music, and an infinite number of filter taps, you get a perfect brick wall filter..

Simon

What I see is claim that the Hugo guess is good one!

If the original recording has a certain time interval [sample rate] between certainties, then clever algorithms can make a good guess about an event in the performance, before the even was recorded, but however good the guess maybe, it remains a guess. Or else we could reconstruct a whole hour performance of a Beethoven Symphony from but half a dozen digital samples!

ATB from George

Source: Hifi Choice

George, really! You would only need the beginning and end sample. Then some guys will tell you it is possible to interpolate the rest with a new computer.

Now, exactly where does the French Horn come in again.....

The brain is a frequency analogue device (similar to variable clock PDM), it doesn't sample as such, and the timings are variable.  The analogy doesn't hold water.

"the brain samples sound in real time every 4 microseconds, whereas CD refreshes its frames every 22 microseconds. Its CD's inability to work as fast as the brain that causes its problems in the time domain, why it doesn't sound natural. And the unique design of the Hugo addresses precisely this failing."

Source: Hifi Choice

"The 4uS threshold implies that the brain samples data at 250kHz – much faster than 44.1kHz of CD recordings"

Source: Chord Website

Correct, I don't agree with it.

They say that Rob said it. Who knows?

A link to this claim would be appreciated.

If you take a look at Robs posts on the subject, he clearly states the opposite.

If you're happy to accept the word of a journalist, that's up to you, but don't expect others to do so.

A link to this claim would be appreciated.

If you take a look at Robs posts on the subject, he clearly states the opposite.

If you're happy to accept the word of a journalist, that's up to you, but don't expect others to do so.