Oversampling: Who Does It Best?

[Miska]

Must be a language barrier thing.

 

Maybe, my native language overloads words less. (same word used less for different meanings)

 

As I type these words in Firefox, both "upsampling" and "resampling" are flagged as being misspelled.

 

Also "datasheet" is shown as misspelled in Opera, so maybe datasheets shouldn't be used as dictionaries. Or place to verify correctness.

 

In the Burr-Brown PCM1792 datasheet, the word "oversampling" occurs 17 different times, both in reference to the 8x oversampling digital filter and also in regard to the delta-sigma modulator section.

 

I (personally) find it messy from their side, some others are a bit more careful on their documentation. AKM says "128x Oversampling" and "8x Digital filter".

 

There's a fundamental difference how these increases technically happen and I find it annoying to use same word to generalize all cases when source and destination sampling rates happen to differ.

 

That's probably a reason why some others wanted to find a different word for things that are fundamentally different. I didn't invent the term, so please don't blame me for it...

 

In any case, I don't think think it is beneficial to get stuck to a word and I'm at least incompetent for being English language police since it's not my native language, and I don't know if I manage to formulate any single English sentence without some kind of language error.

 

[Miska]

I don't really understand the question. I will say that it is foolish to convert to a non-integer multiple of the original data. In this case the ratio is 320/147. It would be far better to covert by 2x or 4x, as the simpler math leads to much smaller rounding errors.

 

It's not foolish, since the beauty is that once you begin to optimize the algorithm you'll notice that there's actually no difference. Thus for me, both take equal amount of processing power and no difference in rounding errors.

 

As a by-product you'll also get the 2x, 4x, etc. optimized.

 

 

[Miska]

I do own a multibit NOS DAC which has NO analog filter, an Audio Note DAC3.1x, and it sounds absolutely fantastic (yeah yeah, they use output transformers to perform that role.)

 

What is needed is "output band-limiting". How it happens is up to the designer to choose.

 

My old Marantz CD-player used it's own interesting type of RLC-filter, even though it had SAA7220 (IIRC) 4x digital filter before it's TDA1541A multibit DAC. Haven't seen coils in these filters that much since...

 

[ziggyzack]

'Please feel free to point out my "inaccurate" statements so that the other readers of this thread can be enlightened.'

 

This would be dull, they can draw their own conclusions if they wish.

 

 

[Miska]

Are there benefits to up/oversample 16/44 materials? and if there are, how best to do it? (letting your software "upsamples" and feed the DAC, or feed them native and let the hardware deal with it etc.)

 

This is naturally my opinion, but yes there is.

 

Let's first declare 20 Hz to 20 kHz as "audio band".

 

Then let's think there are number of cascaded filters. The first one taking rate up from 44.1 has biggest impact on the sound quality, since this filter is closest to the audio band and it's implications spread to the audio band the most. Let's consider for a moment this to be for example 4x filter. The next one taking rate up from 176.4 has significantly less audible impact since it's quite far from audio band. This also applies to decimation/"downsampling" side, when the original recording is made at higher rate than it is being listened. So we could call this "resampling" (?) to cover both cases.

 

Design properties of this particular digital filter largely define how it sounds. There are various weightings and opinions on how these should be designed. And sometimes, especially inside DAC chips there are constraints imposed by cost, heat, power consumption and number of clock cycles available. Modern computer CPU on the other hand can perform billions of 64-bit floating point math operations per second.

 

If it happens to also increase the modulator pulse frequency at conversion stage - even better, since it improves converter resolution. But in many cases modulator frequency would stay the same.

 

Next if we then think about the analog output filter and the work left for it. It has some corner frequency and steepness. Less there is noise in the band that is not filtered or that doesn't yet have a lot of attenuation, the better. So it is beneficial to push the frequency where the noise begins as high as possible and keep the noise before that point as low as possible. This helps reducing noise "dirt" that is still left when coming out of the final filter.

 

What is also related and important is dither and/or noise shaping used after applying such processing and entering the more or less limited sample bit depth of the converter. Newest DAC chips accept 32-bit data input, requiring less "squeezing" of the data before DAC, so there smaller expand-squeeze-expand-squeeze transitions in the processing chain.

 

And in modern world the difference between "hardware" and "software" is not that clear anymore. Some hardware chips have a custom processor and ROM memory inside, running a piece of software (built-in firmware). Some hardware can be defined by software - the hardware configuration is declared by a piece of software (FPGAs). And then there can be a generic signal processor hardware running a piece of ordinary software (DSPs). And then of course just general purpose CPU hardware running a piece of software.

 

[ziggyzack]

My understanding: To get from 16bits to 24bits, you add digits. May be just plain digits or a dithered.

 

I was confused by this initially, so if you are falling in to the same trap as I did, it's perhaps simpler than you imagine.

 

Moving from 16 bits to 24, or to 17 for that matter, doesn't require any scaling of the values, so to speak. You are just adding zeroes. Then, if you want to manipulate the original signal, you have more precision to operate in. No matter how high the precision, because you don't have infinite precision, after manipulating the signal you are still likely to have to approximate the last bit. Rounding is one approximation strategy, but can introduce quantisation distortion as the signal is non-random. Dithering is another - it aims to prevent audible quantisation distortion at the expense of adding impercebtible noise.

 

This is really worth watching, it has demos of quite coarse and easily discernible quantisation distortion, amongst other things. There's an hour of it, although the last part is about video: http://www.xiph.org/video/vid1.shtml

 

In the Linn DS players I'm familiar with, all source signals are up/over/re-sampled to 384/352.8 KHz, 35 bits. Filtering and optionally volume control is handled at these levels prior to feeding the Wolfson DAC. The sample rate is unchanged in the feed to the DAC, while bit precision is reduced to 24.

 

[Charles Hansen]

There are so many postings containing so much inaccurate information that it would be a full time job to try and correct them all. So I will leave with just one example:

 

In the Linn DS players I'm familiar with, all source signals are up/over/re-sampled to 384/352.8 KHz, 35 bits. Filtering and optionally volume control is handled at these levels prior to feeding the Wolfson DAC. The sample rate is unchanged in the feed to the DAC, while bit precision is reduced to 24.

 

In this case the poster (it doesn't really matter who) is simply repeating the BS that Linn's marketing department came up with. But it is dangerous BS, because it fools the consumer into thinking they have something that they don't.

 

Specifically I am speaking of the digital volume control. They say that they oversample (interpolate!) to 8x at 35 bits. This then goes to their volume control.

 

So of course the consumer thinks, "Aha! A 35-bit volume control! This means that I can have a perfect signal from my volume control because I can attenuate by -60 dB (which is the most I will ever need) and still have 26 bits of precision!"

 

And that is what Linn wants you to think. But it is rubbish.

 

The signal from the digital volume control must be fed to the DAC, which only accepts 24-bit input signals. So a high-res source (24-bit) cannot use the volume control at all without incurring a loss of resolution. And a 16-bit source can theoretically accept up to 8 bits (-48 dB) of attenuation, although actual listening tests show that degradation occurs with anywhere between 0 dB and -20 dB of attenuation, depending on the resolution of the rest of the system and the listener's hearing acuity.

 

It's this kind of garbage that makes me ashamed to be a part of this industry. And to think when I founded Ayre 18 years ago, I held up Linn as a role model to emulate. I guess times have changed.

 

[ziggyzack]

In this case the poster (it doesn't really matter who)

 

It was the post just below, seems like you cannot even bear to spit my name!

 

So of course the consumer thinks, "Aha! A 35-bit volume control! This means that I can have a perfect signal from my volume control because I can attenuate by -60 dB (which is the most I will ever need) and still have 26 bits of precision!"

 

Incorrect. I don't think this for a moment. The extra precision allows for multiple processes to be applied to the signal without compounding rounding errors.

 

[barrows]

Charles Hansen is the President, head honcho of Ayre Acoustics, makers of the QB-9 USB DAC, among many fine audio electronics.

Steve is the head honcho of Empirical Audio, maker of the Overdrive DAC, and many advanced computer audio electronics.

This discussion may have strayed too far from being helpful to most audiophiles. I would like to add a little product specific information, and then some advice. If any of this product information is in error, I hope that those who know, will feel free to correct me:

 

Ayre QB-9 DAC: This uses an separate chip (FPGA) running custom digital filters (call it oversampling) and the (oversampled) data is then fed to a BB DAC chip, bypassing the built in filters/oversampling of the DAC chip. This might be termed "single pass" digital filtering (for marketing purposes).

 

 

 

Benchmark DACs: These use an separate, asynchronous, upsampler (marketing term, which usually denotes asynchronous operation) to feed the DAC chip. In this case I am pretty sure the DAC chip still applies another digital filter/oversampling.

 

Berkely Audio DAC: This uses a separate chip, running a custom asynchronous upsampler/filter, which then feeds a DAC chip. In this case I am not sure if the DAC chips onboard oversampling/filter is also used or if all filtering is done in the separate chip.

 

PS Audio PerfectWave DAC: This uses a separate, asynchronous upsampling/filter-this asynchronous filter can be user defeated from the front panel or remote-to feed a Wolfson DAC. The Wolfson DACs onboard oversampling filters are also engaged. So, this DAC can be run with a "single Pass" filter, or with a "double pass" filter, as selected by the user. Additionally, the different filters in the Wolfson DAC chip can be user selected.

 

Empirical Audio Overdrive DAC: As this company does not seem to reveal exactly what DAC chip is onboard, it is unclear to me exactly how it operates, Steve? It does seem clear that this DAC does not use any asynchronous (known as upsampling) digital filtering. This DAC has a number of user selectable filters-I assume these are implemented in the DAC chip itself, or??? Perhaps Steve can shed some additional light on how his DAC operates.

 

My experience has generally been that asynchronous sample rate converters in DACs generally hurt the sound, but I am not sure this is always the case. My recent experience with the ESS 9018 DAC chip (which uses an onboard ASRC in default operation, as used in the Weiss DAC202) makes me wonder a little if some ASRCs are much better than others. In any case, as we can see from the few examples above, there are many approaches to designing DACs, and designers/engineers often have passionate opinions about the best way to go. As audiophiles it is clear we need to listen to the products to decide what we like best. I find the effects of different filters/up/over-samplers to be somewhat subtle, but quite meaningful musically. I would also suggest that longer term listening, with a lot of different music types, will better reveal the differences in digital filters/up/over-sampling than quick A-B comparisons.

 

[Miska]

My experience has generally been that asynchronous sample rate converters in DACs generally hurt the sound, but I am not sure this is always the case.

 

One thing is that it depends if it is used for jitter rejection of jittery source. And in this case heavily depends on parameters of the control-loop.

 

If it is used just for resampling it mostly depends on the algorithm and if it's a "true" ASRC (ie. completely freely on-the-fly ratio controllable).

 

 

[PeterSt]

"Upsampling" is just a marketing term made up by a certain company. It is not an engineering term, and it means whatever the marketing person wants it to mean that day.

 

Of course I took this quote from somewhere when the context was still quite ok (about doing it another time preceeding an existing digtal filter). Still, I see no wrong in upsampling as such, if you only see the analogy with upscaling.

In my book this would be nothing more or less than upscaling the resolution which in the mean time -for audio- acts as a filter (the necessary one). One problem, it is never used like that, unless I refer to my own. So :

 

Any commercially available "upsampler" upscales the resolution allright, but it doesn't better it, because it is a filter in the first place (with ringing and all).

Turn this into an explicit (and interpolating) "upscaler", and the resolution just *is* better, and you'd have the filter in the mean time.

See video upscaling to understand that this really can work, although it will never touch (hires/HD) reality, and the means to do it will be very different.

 

Oversampling ? what about something that samples more than functionally necessary (meaning : for technical reasons only, like making a DSD chip to work (sigma-delta)).

 

Yep. Made it all up. Sorry.

But I see no way a software "upsampler" will call itself an oversampler, while all what upsamples to a certain degree betters the resolution. This incluses the "8x" in the first stage of a chip (environment).

But as implied already, to make it really sound "fitting", upsampling should be about explicitly bettering the resolution (by means of a higher sample rate).

 

And this is the case nowhere.

Ok, 1 (sorry).

 

Peter

 

 

 

 

 

 

 

 

[Miska]

The signal from the digital volume control must be fed to the DAC, which only accepts 24-bit input signals.

 

Well, bunch of recent ones accept 32-bit input. But that's irrelevant, because even at 24-bit the noise floor is so low that it usually disappears in noise of following electronics.

 

And a 16-bit source can theoretically accept up to 8 bits (-48 dB) of attenuation, although actual listening tests show that degradation occurs with anywhere between 0 dB and -20 dB of attenuation, depending on the resolution of the rest of the system and the listener's hearing acuity.

 

This is interesting claim, given that for example BB PCM1794A DAC has only 66-levels in it's conversion stage (roughly 6 bits). Would be interesting to know what kind of algorithm was used in these tests.

 

Now I'm again thinking if I should make available some of the 192/4 (120 dB attenuation) and 192/8 (96 dB attenuation) test files which are pretty extreme cases but very well highlight things.

 

Quality of digital attenuation is quite complex topic and largely depends on implementation details.

 

[ziggyzack]

PS Audio PerfectWave DAC: This uses a separate, asynchronous upsampling/filter-this asynchronous filter can be user defeated from the front panel or remote-to feed a Wolfson DAC. The Wolfson DACs onboard oversampling filters are also engaged. So, this DAC can be run with a "single Pass" filter, or with a "double pass" filter, as selected by the user. Additionally, the different filters in the Wolfson DAC chip can be user selected.

 

For their disc transport or the network bridge, their two preferred sources, PS recommend 'native mode', defeating the initial 'source rate converter' (using their lingo). So all the DAC features for a PWD + bridge are the standard features of the PWD's Wolfson WM8741 chip.

 

[PeterSt]

The signal from the digital volume control must be fed to the DAC, which only accepts 24-bit input signals. So a high-res source (24-bit) cannot use the volume control at all without incurring a loss of resolution. And a 16-bit source can theoretically accept up to 8 bits (-48 dB) of attenuation, although actual listening tests show that degradation occurs with anywhere between 0 dB and -20 dB of attenuation, depending on the resolution of the rest of the system and the listener's hearing acuity.

 

(the bold is mine - and my subject here)

 

I would like to see how these listening tests were conducted.

IOW, I don't concur with it much. Not when a proper digital volume has been used. And not many of those exist ...

What about again 1 (and again sorry).

 

I'm not trying to be funny, but I do know what it takes to get there.

Nevertheless it would be true for a general rule. So, no real problem with the quote above. Still it shouldn't be a "statement" as such, if it's only done right.

The other way around, when such a test would be conducted with any random (other :-) digital volume, there should be differences perceived, if only -indeed- the system is resolving enough.

 

Just putting in some counterweight ...

Peter

 

[PeterSt]

The word length is increased by any mathematical operation, usually multiplication. If you take a 16-bit number (such as the data on a CD) and multiply it by a 24-bit number (such as the coefficient in a digital filter), the result is a 16 + 24 = 40 bit number.

 

The problem is that the 40 bit number must be shortened, because there is no such thing as a 40-bit DAC chip. So the smaller part of the number (LSBs = Least Significant Bits) are discarded, either by truncation, rounding, or dithering. The goal is to return to something that the DAC chip will accept, such as 24 bits.

 

The latter part would be true allright - and in the context of the post this was in, indeed this does not imply 40 bits resolution anywhere (but please keep in mind that *these* 40 bits were a subject in the mind of the poster, and I don't think someone else's). And then :

 

The first part is beyond me. Maybe it assumes some standard DSP chip operating in the 40 bit domain - I don't know, but it is nothing for any normal - or if you want decent operation.

 

The text in the first part - and merely the result shown - will be true allright, but the whole operation of some necessity (??) of multiplying a 16 bit word with a 24 bit word is not. Use 8 bits - done. Or use 16 bits for 32 bit output words.

 

No further subject of truncation - dithering or anything is there now.

 

Why it was put like how it was put ? I don't see it at this moment. But I don't know all either.

Peter

 

[PeterSt]

Maybe it now starts to look like nit picking, but indeed so many has been said in this thread which is more or less off, that maybe I had to put my focus somewhere ...

 

The digital filters in most DACs are pretty good and adding another digital filter in front of that is not going to help things, unless for some reason it is unusually good.

 

Maybe it depends on one's interpretation of "pretty good". But I have never been amongst that group who thinks this. Merely the other way around : all filters (and they are all about equal) s*ck as hell because the principle ain't right. Ok, says me. Also, the "group" I'm in is pretty small, because I'm quite alone. Maybe not for listening experience, but for thinking about the design principles.

 

The above is totally unimportant, but assumed true (what I obviously do), something else isn't right now :

 

and adding another digital filter in front of that is not going to help things

 

See ? now this one doesn't fit much anymore. In my thinking any random shot at creating a (software) filter in front of existing may help.

But of course, you added "unless it is unusually good", which makes the text not much wrong again.

 

But there's the suggestion that this can't be anyway, or at least it won't be good ever to "concatenate" filters. And again, true. However (and this is as far as my thinking can go at this moment), I think it is good to understand for people that once an earlier "filter" has done its job in a better way than a next one in the chain, that next one isn't "capable" of destroying what the first did. It's all maths and harder to explain (by me), but also empirical finding by the largest group of ABX you can imagine. I mean :

 

That better filter does exist (not important by itself), and it prooves that it works like I described, because almost without exception everybody uses it. This (for me - back then) unexpected result is almost worse, if you know I advised people *not* to use it, just because it indeed would be that other filter in the chain. Still they do, and they do always.

And you can well say that from there I learned how this really works.

 

Ok, I'm only telling indirectly that I don't know much about it all (to some degree, ok ?), but the message - *and* the subject is : software "filters" for 100% sure have a fair chance. It only needs taking distance from being that similar filter which is in the DAC already (which will only destroy because of inconsistencies between the both), and do a decent job instead.

 

While my text here is hardly a suggestion as such, I thought it could be important to at least suggest the complete other way around from the quote this post starts with. Maybe without agreement between respective posters, but still something to think about more in the future.

 

Peter

 

[PeterSt]

Allow me, please, to paste in just one other block of text, maybe by now not without coincidence from again the same poster. Maybe I'm intrigued, maybe it is out of respect, maybe it sets my own mind ... but where this indeed is just a copy paste from one adjacent piece of text, so many things can be interpreted differently from, well, how I do. Don't be offended, ok ?

(and indeed think "we" could go on and on and on in this thread, but I really will stop after this one)

 

The below may be seen as out of context to the original texts, although the context is copied in the quote(s) itself to a certain degree. But try to read it as how interpretation could be.

 

One poster claimed that non-oversampling sounded far better than using any kind of digital filter, and that NOS DACs were very popular. There are three problems with this:

 

1.

a) NOS DACs are not very popular. There are a handful of companies that make them compared to probably a hundred that use digital filters. There is a reason for this.

 

2.

b) A few posts later, the same person said that "upsampling" would improve the sound. This is nonsense. An "upsampler" is a digital filter. So a digital filter both improves and ruins the sound??? Please explain this one to me.

 

3.

c) We spent four months listening to digital filters of all types, including non-oversampling (NOS). NOS has its charms, particularly in the mid-band. But it also has its flaws, particularly in the frequency extremes.

 

Ad 1.

 

I don't agree with this. If we were able to dive into this really (and I think I can), we'd see that NOS DACs are used by people with the finest systems. Maybe not always the $$$,$$$ systems, but often so anyway, and always about having the best sound possible. This is different from getting yourself a 5K+ DAC and connect it to your system, sit back and have a nice sound. Of course, the latter group can behave exactly the same as the former, but there's a trend which makes me say that NOS is more popular than OS, were it for the quest for the best.

 

No agreement on this ? ok. But this is different from pointing to a smaller group of manufacturers creating NOS, and derive it from there. It only prooves that manufactures may not get it. Or, that it is way more difficult (thus expensive) to create well performing NOS (my very own "statement").

 

Lastly on this one : the NOS using group may be a 100 times or more larger than the official "NOS manufacturing" group, if you count in the DIY group, which from logic (no manufacturers) create a way of their own. And sure not because it would be more easy.

 

Maybe the latter is a stupid argument, but let's say it belongs to the perspective.

 

Ad 2.

 

Not talking about the both improving and destroying at the same time, an upsampler *can* improve the sound;

 

Apart from your own suggestion it can (see previous post - and this is not about word games, but about knowing what we both talk about and actually agree upon) - it is done. But please keep in mind : only when this is about "upscaling" as how I present it.

 

Btw, although not originally I think, I now put this in the context of NOS, which is your own "stream" regarding the quoted text.

So, for NOS this is of major importance. And yes, a bit pitty that NOS inherently is without upsampling or how we call it, but there is a major difference when we allow ourselves to apply a software filter in front of real NOS which at least will eliminate cascaded filters (which *is* a problem, never mind my previous post).

 

Pieces of the puzzle might come together here ...

 

Ad 3.

 

It depends a bit on wording I guess, but once a "filter" is NOS itself (read the quote), I'm not sure how things are compared. So, I don't see how NOS as such is a filter itself, if only it's NOS without a(ny) filter. Oops.

 

In order to again let your texts be allright and all over, I will admit that any commercial NOS filterless DAC which is 16 bits won't be able to go anywhere, were it for listening or good figures. It just can't. The 18 bits have a better chance, but only because of some digital means of filtering now is allowed (possible).

The real deal though is 24 bit, and while ever one existed that I know of (MSB) and while nobody seems to know to what degree the legendary PM-II is really filterless or not - only such DACs allow for ...

 

Decent filters.

 

Well, I don't recall I have ever been more confusing, but this one is on purpose, and a bit for fun maybe. The whole point is though : only DACs without any means of filtering allow optimally for software filtering. And wasn't this thread all about that ? In other words, any DAC containing means of filters are lost for this to begin with (not judgeing whether they are better or not), which is what I think you and me both say in the end for the general rule.

 

So ... NOS is not a lost case at all (just my words) but instead gives opportunities not there before. And this is the software (pre) filtering, or "upscaling" to one's creativity.

And in case it is not clear : Only real NOS/filterless DACs allow to judge the merits of that means of filtering. Other cases will present a mixed bag of unpredicted results.

 

I hope we can agree on this, and otherwise it's my vision anyway.

Peter

 

[barrows]

Still an interesting discussion. What I am drawing from this is that opinion varies quite widely on what is the best approach to produce the best sound from digital files, and I would not for an instant disagree with the listening experiences talked about here. I know that at least Peter and Charlie's experience is based on listening to the their approaches with music.

Peter: my understanding is that your approach is using a NOS DAC based on the TI (BB) 1704, but your approach is not truly what most would call NOS, as you do oversample in the computer, via your own, unique and sophisticated, code.

To be clear: where Charles Hansen noted that the filters in DAC chips are "pretty good", I think a bit of language barrier might be at work for those reading. "Pretty good", does not denote that these filters are as good as they can be. Remember, that Charles' company, Ayre, has gone to great lengths to design their own proprietary digital filters, and does not use the on board filters in the DAC chip-so it is clear that while Charles calls the DAC chips' filters "pretty good", he certainly believes they can be improved upon.

Miska: I am pretty clear on Charles' and Peter's approach to digital, and what they prefer to listen with, and some of their background. I am not clear on yours. I am sure you have avoided marketing any products directly here in respect to the forum, but I would love to hear what your preferred approach is in terms of listening to digital music files, to add additional context to this discussion.

 

[Charles Hansen]

I forgot to unsubscribe to this thread, and was compelled to read it like watching the aftermath of a gruesome automobile crash.

 

Most of the posts were either unintelligible or else claimed my posts were wrong but didn't bother to say way they felt was right. However, one post by Barrows was quite good and tried to present a balanced picture of various perspectives. There was one detail I wanted to clarify, because I both understood the post and it also referred to our products.

 

Specifically, what Barrows referred to as "single pass" versus "double pass" digital filtering. Barrows assumed that single pass meant that the digital signal only went through one digital filtering chip, while "double pass" meant that the digital signal went through two digital filtering chips (for example an external digital filter that feeds a DAC chip with an internal digital filter).

 

When I use the term "single pass", I use a much more strict definition -- literally a single pass through a single-stage digital filter. Allow me to clarify.

 

By far the most common way to perform high oversampling rates (anything above 2x, such as the 8x found in most DAC chips) is to use a concatenation of 2x filters. For example, it requires three 2x filters to make an 8x digital filter. The reason for doing this is because it is far cheaper. It typically requires less than 1/4 of the computational power (read: silicon area) to do it this way.

 

The problem is that at each stage the output of the digital filter is too long and must be dithered, rounded, or truncated before sending it to the next stage. These multiple rounding operations lead to a cumulative error. In contrast, in the Ayre products the digital filter is literally done all in one big step.

 

And that is what I mean by "single pass" -- it is literally a single pass.

 

To my way of thinking, what Barrows described as "double pass" (when there is an external digital filter chip feeding the internal digital filter inside the DAC chip) is meaningless. So now there are three stages of filtering inside the chip. Adding another stage (or even two) outside the chip doesn't change things as you still have multiple rounding operations.

 

This is also why I don't like to use IIR (Infinite Impulse Response) filters in our products. An IIR feeds part of the digital signal from the output of a digital filter back into the input of itself. It must be rounded, dithered, or truncated. Then it makes many, many passes as it recirculates (not an "infinite" number, but that is where it gets its name). But with each pass the rounding errors accumulate.

 

And to be historically accurate, we were not the first to use single-pass digital filtering. Oddly enough, it came from companies like Sony and Pioneer. Most of these models were super-high-end designs that were only sold in Japan. But a few made it to the States (and Europe). Probably the most widely known is the Sony SCD-1, the very first SACD player. If you still have the literature, go back and look at a chip they called "VC-24". It performed 8x oversampling in a single pass.

 

[Charles Hansen]

Peter ST wrote:

 

The text in the first part - and merely the result shown - will be true allright, but the whole operation of some necessity (??) of multiplying a 16 bit word with a 24 bit word is not. Use 8 bits - done. Or use 16 bits for 32 bit output words.

 

That seems like an easy solution.

 

But it doesn't work.

 

Try it sometime and you will see. If you use 8-bit coefficient with a digital filter for CD (16-bit data), the output will be 24 bits. But it sounds like hell, because the effective resolution is actually only 8 bits.

 

[Miska]

Miska: I am pretty clear on Charles' and Peter's approach to digital, and what they prefer to listen with, and some of their background. I am not clear on yours. I am sure you have avoided marketing any products directly here in respect to the forum, but I would love to hear what your preferred approach is in terms of listening to digital music files, to add additional context to this discussion.

 

My background is mostly in custom hardware and software for all kinds of signals, audio and others.

 

On music reproduction, generally my attempt has been to extract best possible performance out of the hardware I have. And I have various different pieces of it. Nothing really expensive though. I mostly ask myself how good sound is it possible to get out of particular reasonably priced or inexpensive piece of hardware by aiding it with software. My second question is usually how much hardware can I replace with software. I'm not afraid of replacing analog things with digital ones when there is a technical justification for it. But I don't like to do it for the sake of being digital either.

 

Usually any particular technology has it's sides and it is important to understand these.

 

I also have to admit that I'm hopeless perfectionist, never completely happy with anything. I don't know any piece of hardware on the market which would be exactly the way I like. Usually one piece has one part of it done the way I consider correct while some other part is not. That's why I have my own designs too.

 

Since I tend to work when I'm not sleeping, I mostly listen using headphones while working. So the ratio is roughly 75% vs 25%. And I also strongly believe that given same amount of money, it is possible to get better sound with headphones than with loudspeakers (and I'm talking about four figure or less price tags here).

 

I also demand that things both sound good AND measure well. As well as being as "technically correct" as possible.

 

I'm also one of those who believe that CD is technically insufficient and requires some serious DSP to make anything proper out of it. That 176.4/24 is the minimum good one and that LP, SACD and DSD128 are great things... I could also spend ages explaining the technical details why I think so (maybe I've done too much of it on this forum already).

 

Now I don't know if I answered the question at all, but I tried...

 

Edit: For those who are interested in what I have mainly in use.. My amp has 2x AD1955, my own DAC design has CS4398 and EMU 0404 USB has AK4396...

 

 

[Miska]

The problem is that at each stage the output of the digital filter is too long and must be dithered, rounded, or truncated before sending it to the next stage. These multiple rounding operations lead to a cumulative error. In contrast, in the Ayre products the digital filter is literally done all in one big step.

 

For those interested in this technical detail, all the "poly-sinc*" resamplers in HQPlayer are single-pass and synchronous, between any of the supported standard audio sampling rates, upwards or downwards in sampling rate. Including for example 44.1 -> 96 and such.

 

 

[PeterSt]

Peter :

 

The text in the first part - and merely the result shown - will be true allright, but the whole operation of some necessity (??) of multiplying a 16 bit word with a 24 bit word is not. Use 8 bits - done. Or use 16 bits for 32 bit output words.

 

Charles :

 

That seems like an easy solution.

 

But it doesn't work.

 

Try it sometime and you will see. If you use 8-bit coefficient with a digital filter for CD (16-bit data), the output will be 24 bits. But it sounds like hell, because the effective resolution is actually only 8 bits.

 

Ahh, I misunderstood (or did not get) that this included the filtering "stage" itself. So, of course.

 

(notice that one can also create the *base* for the filtering by this multiplication, and next interpolate or apply other means to fill in the gaps)

 

 

[PeterSt]

Peter: my understanding is that your approach is using a NOS DAC based on the TI (BB) 1704, but your approach is not truly what most would call NOS, as you do oversample in the computer, via your own, unique and sophisticated, code.

 

Correct (but what did you say there ? OVERsample ?). Still the NOS1 is fully NOS/Filterless of course. So, it would allow people to get the true merits of this - which may not be *that* special within itself, but there's much more which actually works out counter productive so to say. Think like near to zero smearing because of very low jitter, and next it sounds completely wrong because it puts out what's in the data 1:1 (I have some recent graphs of that somewhere). So, "smearing" (by whatever means) would act somewhat as a filter we talk about all the time, but this doesn't happen, and so it shows how bad NOS/filterless really is. Maybe sounds strange to you, bit it really works like that.

 

Next though, we are free as can be to create that filter in front of it. Try them all. Compare. No intervenence from something inside.

 

That there's also nothing to interfere inside at 96, 192 or 384 input sample rates is another thing. But the NOS1 really is not made for that. Oh, it is as good there, but I hunt for redbook to be as good as possible. There's so much more of that ...

 

Your observations remain correct (when are/were they not ? haha). But think about the possibilities. This is not only about using all the filters (and their settings !) you can find in the world - which are not hindered by any means inside. It goes beyond that, but I guess you must have followed the development a bit to understand. And I'm not here to plug it either, so never mind for now. The point for here is the filtering, and indeed the software in front of it which may make something nicely useable of that because it should work in real time. XXHighEnd does that, and e.g. HQPlayer may do that even more extensively.

 

It makes the sound ... (trust me)

 

Peter

 

 

[ziggyzack]

'Most of the posts were either unintelligible or else claimed my posts were wrong but didn't bother to say way (sic) they felt was right.'

 

Well, I think I corrected you pretty clearly here, when you tried to put words into my mouth:

 

Incorrect. I don't think this for a moment. The extra precision allows for multiple processes to be applied to the signal without compounding rounding errors.

 

Funnily enough, after I talked about the issue of compounded rounding errors, then so did you. Perhaps if you weren't so guided by picking holes in other products/people, you would have got there directly from the 1 page (so clearly marketing, but not misleading) Linn DS upsampling document.