English Premier or International (HiFi) - Now with Team History

[Simon-in-Suffolk]

Now - this got me thumbing through some of my old text books to get the details right rather than rely on my memory…

In the real world it is effectively impossible to provide a pure pulse to gain a filter response from. Pure pulses have zero time! So most DACs use reconstruction filters that are designed to responded to a zeroth-order-hold - ie the sample value remains consistent until the new value is received. This requires the spectrum of the sample stream to be multiplied with a sinc function to provide the equivalent spectrum response of a pure pulse stream.

This actually means the higher frequencies in the spectrum are increased in amplitude up to the cut off frequency if the spectral response is to be accurate in the reconstruction filter .

So we then get to the low pass cut off filter itself in the reconstruction filter - and in audio this is often a Butterworth filter as opposed to say a Chebyshev or Elliptic filter - as the latter have a sharper cut off but introduce spectrum ripples in the pass band. The Butterworth filter has no pass band ripples, but normally requires added poles to increase the cut off rate so as to avoid aliasing. The Butterworth filter also has a more linear phase response through the pass band. Over sampling helps in this regard too by increasing the sample rate with respect to the actual pass band thereby reducing the likelihood of aliasing for given filter slope or order.

So in practice a so called ‘brick wall’ to my mind is a term that is not helpful. It’s the order and type of filter, and the sample rate with respect to the pass band that is important.

So back to your view of rolling off the high frequencies in the reconstruction filter - well as I said this happens anyway for zeroth-order-hold DACs, and would normally be compensated for.

However generally rolling off the higher frequencies might stop other issues on lesser equipment such as intermodulation distortion and other artefacts in the analogue chain and so might be preferable - effectively you are applying a tone control by distorting the reconstruction filter.

That is how I understand it from my DSP days.

I am told in some more basic DAC and filter setups - the reconstruction sinc compensation is sometimes omitted altogether.

[Edmund-of-Essex]

Hi I_B,

Thanks for considering my question and your quick summary response.
Really, I am at a fundamental decision point on my various projects, summarised thus:

1. Can I be bothered to find out what the effective impulse response is of the Linn Organik DAC?
2. Am I going to gain any more useful knowledge re selecting the ‘correct’ loudspeaker cable for my system (beyond the measurements that I can already obtain using my existing MiniDSP Umik-2 with REW employing the highest sample frequency available - 192 kHz)?

One professional audio device that had caught my attention is the RME ADI-2 Pro FS R.
A better approach might be to buy the correct piece(s) of laboratory bench test equipment(s) - but then things get exorbitantly expensive - and then what is the ROI?

Thanks for your thoughts.

E of E

[Innocent_Bystander]

Leaving you to ponder your first question, on the second I think if it were me I’d stick with 192, the is the tests don’t show any difference I’d make cable decisions purely on sound preference, though the reasons for those sound preferences would sadly remain unknown, but what might sampling at 84 capture that 192 doesn’t? Rhetorical question, however I think it would be in the realms of the finest subtlety not major differences, just as when comparing hi res files with 16/44 from the same master.

And you would still have the relationship, if any, between cable parameters and any audible differences in sound - which if there is a correlation is very valuable information. And although some aspects he differences in sound may be dependant on the amp and speaker the cables are connecting, the results may still be indicative of possible direction of sound change between different cables.

[Edmund-of-Essex]

Hi Simon,

Thanks for consulting your textbooks. Of course in the real world (S domain) it is impossible to provide a pure pulse. However in the Z domain - sampled/digital representations of ‘signals’ - it is possible. In this case a digital ‘pulse’ can be created by lots of zero values followed by a single sample with an equivalent value of ‘1’, followed by lots of zeros…

The key question then becomes, what mathematical technique(s) is/are applied (and with what practical implementation details) to get back into the continuous time reality (S domain) when reconstructing and original analogue (continuous time domain signal)?

A quick summary for the benefit of the discussion and any Naim forum readers who might not have come across terms and descriptions like this before (the Convolution Theorem):-

Multiplication (of signals/waveforms) in the Time Domain is equivalent to Convolution (of spectra of the signals/waveforms) in the Frequency Domain.

Multiplication (of the spectra of signals/waveforms) in the Frequency Domain is equivalent to Convolution (of the signals/waveforms) in the Time Domain.

I think it is worth clarifying something WRT your observation:-

@Simon-in-Suffolk

So in practice a so called ‘brick wall’ to my mind is a term that is not helpful. It’s the order and type of filter, and the sample rate with respect to the pass band that is important.

I think I have not quite properly communicated the point I was making. IMHO the Information Theoretic viewpoint of ‘all information can be retained up-to FS/2’ has caused so many problems for digital music record and replay at FS of 44.1kHz. This is because (as you have noted) the filters (whether digital or analogue) to prevent aliasing at the point of recording, and similarly the reconstruction filters for play back at the DAC, have necessarily had to trade achieving excellent stopband rejection by using high order filters (ideally with no passband ripple - e.g. Butterworth) at the expense of extended transient and/or ‘pre echo’ behaviour in the corresponding impulse response.

The point I was trying to make is that some modern DACs (and even some bespoke products such as dCS) facilitate designer or even customer choice regarding these trade-offs.

For example the ESS device in the Naim CI-102 offers eight different filters. Here are the impulse responses of the first two options of impulse response:

Credit: ESS Technology Inc. Reproduced for fair use illustration of various impluse responses.

Here are the impulse responses of the final three options of impulse response:

Credit: ESS Technology Inc. Reproduced for fair use illustration of various impluse responses.

These images demonstrate the amount of difference between the impulse responses and it is these functions that are Convolved with the incoming recorded music samples data stream to begin to recreate the output analogue waveforms.

What does the corresponding frequency domain characteristic look like?
Here are the frequency responses corresponding to the first two impulse responses shown before:

Credit: ESS Technology Inc. Reproduced for fair use illustration of various frequency responses.

Here are the frequency responses corresponding to the final three impulse responses:

Credit: ESS Technology Inc. Reproduced for fair use illustration of various frequency responses.

If one looks carefully at the final two filters on the last picture, it becomes obvious the filter roll off starts earlier than 20kHz (and is quite unlike a ‘brick wall’). I know that is lazy language - nevertheless it creates a ‘picture’ in the readers mind - and for these ‘soft edge’ filters, the corresponding impluse responses are shorter in time.

Note that many of the impulse responses have impact out to 1 millisecond and will be affecting the phase of the output signal as well (as they are ‘causal’ in implementation - with the exception of the Linear Phase Apodising filter (no 2)).

In my view, these are not ‘distortions’. The ‘Information Theoretic’ view point of Shannons sampling FS/2 has lost the reality of the sampled music waveform is already distorted as music is not a deterministic signal - it is at best stochastic and has components beyond the FS/2 value of 22.05kHz at FS of 44.1 kHz. FYI, I have personal engineering knowledge regarding this - I have always selected an FS of 4 times the Shannon limit (for interpolation and decimation) specifically to preserve transient behaviour performance in any subsequent processing and reconstruction filters for analogue output/replay.

Now, my question is:- which filter (from the choices in the ESS device) did Naim pick for the CI-102?

[Simon-in-Suffolk]

Yes but you are describing the filter responses that I was referring to. The filter response is typically modelled on the Butterworth filter with added poles, this eliminates ripple in the pass band, and the ripple rapidly goes to zero in the stop band. Added poles creates a steeper roll off.

For accurate spectrum reconstruction with respect to the sample stream the sinc function is the mathematical function that is applied to filter response to map a zeroth order sample stream to a sample pulse stream response for a given filter. That is the sinc(x) . F(x) where F is the filter function. Sinc(x) is sine(x)/x.

Other tone control type filtering effects can be applied to reconstruction to suit taste, but to my mind, that is separate from actual sampling reconstruction.

Naim adopt for reconstruction a digital Butterworth low pass filter with 6 poles to provide the required roll off for their design with respect to Nyquist frequency aliasing.(with their ratios of pass band and over sampling)

As far as the actual digital implementation Naim use an IIR (Infinite impulse response) of the 6 order digital Butterworth algorithm as opposed to FIR (Finite Impulse Response) of the 6 order Butterworth filter using the the programmable functionality of the Analog Devices DSP SHARC processor. IIR uses algorithmic digital recursion, where as FIR uses a defined filter response over a defined or finite length of samples (taps) which is convolved with the sample stream values.

Naim publish various product design white papers that describe some of their design choices, and show this approach has been consistent since the first dedicated Naim DAC product.

[Edmund-of-Essex]

Hi Simon,

Thanks for your response - I feel our technical conversation is heading towards some common ground. So my reply is …yes, I am describing the filter responses you were referring to and yes I do know why the Sinc(x), (cf SIn(x)/x) function is needed (and incorporated) as well.

However, one of the things I was trying to do was to engage you in a conversation that might move away from text book discussion to engineering reality. You have made reference to the Naim approach of using an IIR (Infinite impulse response) implementation of a 6th order digital Butterworth algorithm. That is fine - I have read and fully understand the Naim white paper on the topic and the reasons for the various design decisions.

However on a point of clarity, the computational load and memory requirements of implementing general form Butterworth in DSP as an FIR are not worth it. Just by ‘inspection’ one can conclude that. In the Naim white paper there is discussion about floating point arithmetic and the benefit of 40 bit FP processing in the selected DSP device (from the Analog Devices SHARC family).

Nevertheless, IMO, there are some significant missing comments in that Naim white paper. No discussion for example about the alternative and beneficial use of Fixed Point processing combined with convergent rounding (i.e. when and where to apply that in any code that implements IIR or FIR filters). But I guess that Naim were swayed by the ease of software development promised by the Analog Devices SHARC family (I might have chosen differently).

Regarding your last paragraph. That is a bit more complex. I am not sure that Naim still use this code approach for the new products using the ESS 9033Q DAC chips, i.e. CI-102?
Perhaps they do retain this approach - after all the code is already written.

However, if the selected DAC chips have superior choices of impulse responses - and by implication better sound than a straightforward pole adjusted 6th order Butterworth - why wouldn’t Naim drop the IIR Butterworth code, use what is available in the DAC and instead concentrate on developing underlying DSP code to support Focal Speaker correction and Room EQ code and ADAPT software?

PS: I got zero response to my question regarding this specific point to Stevesky when I was completing my assessment of the CI 102 (CI 102 Review Question to Naim). FYI I find the radio silence from Naim odd. After all it would have been easy for someone at Naim to say - it is as per the NAIM white paper on DACs (2009)?

[Simon-in-Suffolk]

ok a bit to unpack in that one..

With regard to FIR implementations, historically the computation power was indeed high as you say for large filter kernels such that any benefits in using that approach were generally negated by the increase processing noise convolving the Butterworh with sinc kernel response.. now with modern advanced low noise and low power electronics such as with certain modern FPGAs the computation can be practically undertaken with minimal side effects.. and many DAC manufacturers do indeed this now - including Chord Electronics with their very large kernel filter responses (taps)

As far as whether to use inbuilt filters or separate filters - and indeed the PCM1704K has inbuilt filters (albeit bandwidth limited) that Naim disable. Why? I suspect this is all down to evaluation of what sounds best. Now I think I am correct that the PCM1704K was limited to 96kHz oversample sample rate with its inbuilt filter - and that was not enough for Naim.

We shall see with later DAC chips what Naim do - though I suspect they will filter separately to meet their sonic requirements.

I think the jury is out on hifi room correction - I have yet to hear a really good top performance composite implementation in my opinion - and that includes from Linn. I think it’s great for home theatre etc - but for fine music ? - for many of us I suspect we would find ultimately detracting… there are, I guess, too many compromises in standard loudspeaker constructions.

I use the word ‘composite’ because I have heard a good performance where there is a band pass eq DSP for each active driver in an array.. but these tend to be rather expensive and more specialised.. I guess these are more like configurable cross overs for higher power speakers.

I have heard the Naim Bentley sound systems too where they use active DSP to adjust the sound - all very good and impressive - but not really immersive listening quality (in my opinion) - but probably just as well if you are driving!!!

Interesting if you are into this - it might be worth listening to some of the highend B&O designs

[Protegimus]

Hi Edmund,
Couple of potential resources that may be of interest:
DeltaWave Audio Null Comparator

and
QuantAsylum QA403 Audio Analyzer

Have you also been assessing the DAC while conducting your other experiments and tests?

May I ask for the source ESS reference information (I’ve not located the same yet amongst their datasheets, albeit with limited search time, although they do have good text filter descriptions) and what audio interface you use (especially together with REW) for measurements?
In general, this is quite descriptive of effect:

[Edmund-of-Essex]

Hi Protegimus,

Very many thanks for your post and the information summary provided. I think that the table and descriptions you have posted reinforces the point I was discussing with SiS.

Also, thank you for suggesting the software and the audio analyser equipment. Regarding your question about DAC testing. I have captured some data regarding DAC differences (CI-102 vs Linn Organik), but I did not save all of the time domain information for posting as I am still developing my test suite which is based on REW and the MiniDSP UMIK-2.

For DAC assessment using REW, I run at the very highest sample rate available via that microphone and look closely at the ‘edge’ corners of the time domain waveform when a digital square wave (or digital pulse wave) is output as the test signal from REW. One just has to be super careful with edge rates and signal levels as you can blast the tweeters doing this kind of testing!

Re your remaining question - the ESS information was a bit tricky obtain. The ESS website (where the information is available) relies on registering with the website.

However - I am in the UK - and if one uses various electronic parts distributors websites, generally this kind of information (specification sheets) is readily available - so that is what I did. (There are quite a few electronic few parts distributors, the one that I used for the search in UK is Mouser Electronics).

So, if you add ‘Mouser Electronics’ into a search string with the ESS DAC part number you might find the information I used for my post.

BTW, there can be multiple document issues (versions) of a given specification (e.g. v0.4.0 or v0.5.0) available and I have discovered that although the mechanical, electrical parameters and solder flow profile recommendations of a given device are the same between document versions, sometimes the graphs differ in the level of detail provided. FYI I found at least two document issues for the ESS DAC family that NAIM have selected for the CI-102 product. The frequency curve characteristics in one of those document variants was simpler than the other. FYI - I provided the simpler version.

[Edmund-of-Essex]

Hi I_B,

Thanks for this input - much appreciated. FYI I decided to have a go taking some measurements today with a number of loudspeaker cables that I either own or have managed to borrow from one or two HiFi dealers.

The ‘using your ears’ approach is clearly the quickest and most effective way to make a purchase. However I am fairly confident that you, like me, would be concerned with any potential biases in that kind of listening (expectation, order of listening, etc, etc). For example, it was very hard for me to say one way or the other which cable I preferred when Mrs E of E was changing the cable in the preliminaries for World Cup Group A Qualifying tournament (link to previous post).

Regarding the choice of sample frequency, I decided to go with the highest available (i.e. 192kHz), simply for the reason that this setting has the best chance of capturing any weird/unwanted high frequency transients that the amplifier-cable-loudspeaker subsystem might produce during operation.

In my case, the -3dB bandwidth of the Linn amplifier is 120kHz (when driving into 2 ohms), so really 192kHz sample frequency is too low - hence my original question for equipment working up to bandwidths of 384kHz, which would mean sampling at 768kHz.

As an aside, the ‘Megatest of 32 Loudspeaker Cables’ undertaken by Alpha Audio in 2024 was using bench top test equipment with the stimulus and observation bandwidths in the order of tens of MHz!

[Edmund-of-Essex]

Hi Simon,

Thanks for continuing to engage with this topic.

I’ll jump in with the observations on room correction systems for music replay. I agree - it is very tricky to get right. Even the Linn Products Ltd Space Optimisation (SO) approach has many problems and in my opinion the problems have not been sufficiently addressed.

For example, in addition to the additional tuning that may be necessary (see E of E Method to ‘Super Tune’ Linn SO), I believe - but I don’t yet have specific measurement evidence - that the word sizes and computational noise in implementing the Space Optimisation (room correction) filters can prematurely truncate the reverberation decays of instruments. Furthermore, this is for instruments and associated sounds that might generally be considered to be beyond the declared bandwidth of impact of Space Optimisation.

The specific ‘ears’ based observation I have is that in a friend’s room (in 2020 using the SO algorithm of that era) my test piece of music (Deutsche Grammephon, James Levine conducting Wiener Philharmonika playing Smetena: Ma Vlast, track 2 Vitava) had the ‘triangle’ (I can’t remember the exact timing point) initially decays correctly and then just disappears very suddenly (not naturally) when Space Optimisation is turned on. With Space Optimisation turned OFF - one hears the room modes - however the triangle instrument sound decays completely naturally.

I believe this is caused by the (multiple) very deep nulls and high Q filters that SO may ‘design’ to implement the room acoustic correction. The consequences of this are that extra digital signal processing ‘headroom’ will be required in the internal ‘registers’ of whatever digital filter architecture Linn are implementing in the FPGA devices that they have chosen. Luckily in my case - because I have got acoustic treatment as well - the nulls demanded by SO to correct my media room are low in Q and low in magnitude of gain adjustment, hence that potential digital processing dynamic headroom issue/problem is avoided.

Regarding your observations on DSP implementations of crossovers for loudspeaker arrays or each active driver in a loudspeaker array. I guess there are a few high end products using such approaches. B&O as you mention (BeoLab 90?), Linn Products Ltd also have a product (Linn 360) as well as Focal/Naim (the Diva Utopia and Diva Mezza Utopia). FYI, I have not heard any of these - and in any case I have already bought my Magico M2s. Despite what I have written elsewhere on the forum, (E of E views on Active Loudspeakers) what defines a great loudspeaker for me is the dynamic acoustic linearity during normal and elevated music replay. That is why I choose Magico - even though they are so very expensive.

PS: This is also one of the reasons why I have not (yet) bought any fancy loudspeaker cables or interconnect cables!
Nevertheless if Magico launched an active DSP crossover loudspeaker I may be interested in changing?

PSS and BTW: I was somewhat surprised by something you wrote - and it caused me to pause and think: Did I write somewhere on the forum that the Burr Brown PCM1704U has internal up-sampling filters? Because AFAIK it does not - and never has.

The special idea in the implementation of the Burr Brown PCM1704 is/was that it supported digital word input sample rates up to 768k samples per second. This rate facilitated the application developer to use external digital interpolation and have soft (low order) analogue output filters with good transient response. Furthermore the device used sign-magnitude word formats and corresponding internal circuitry to provide excellent linearity all the way down to very low (digital) input signal magnitudes. The PCM1792 DAC (used in other Naim products) however does have such (*8 oversampling) digital filter incorporated within the device. It is maybe this device and these products that you were thinking about?

[Simon-in-Suffolk]

@Edmund-of-Essex

Burr Brown PCM1704U has internal up-sampling filters? Because AFAIK it does not - and never has

Yes it does support oversampling you can easily obtain its technical sheet and you can see the details. It’s an 8x oversampling DAC when used with an external interpolation filter upto 96 kHz with a maximum BCLK of 25 MHz. The actual physical digital and analogue filtering components are external to the chip itself, such as the matching DF1704 chip.

[Edmund-of-Essex]

A quick update………

A few posts (and weeks) ago I said I had not heard the Linn 360 series of active loudspeakers. Well now I have……..and I won’t be changing from my Magico’s.

The Linn 360s were good - but not that good. I think one of the main limitations is that the casing stores too much energy and releases it via the sidewalls of the Linn 360 loudspeaker.

Also during the last three weeks (and in a couple of different locations) I have heard some sources other than Linn DSMs. These other sources were very good and coincidentally neither of which I had heard before.

One of these sources is quite well known on the forum - it is a Chord DAVE.

Elsewhere I reported that I have moved my loudspeakers and listening position by a small amount and as a result I may be able to get away with not using the Linn Space Optimisation feature in my Linn DSM. This creates the potentially opportunity of using other/better DAC sources….such as a Chord DAVE.

I hope to be able to have a home trial soon.

[Innocent_Bystander]

I had a home trial of Dave, when I tried to convince myself it I didn’t want to spend the money. I was right that I dIdn’t want to spend the money, but I failed to persuade myself that buying Dave wasn’t absolutely the right thing to do! I’m stunned and amazed to realise that was nine years ago - since when I have had no interest whatsoever in even thinking about other DACs.

[Edmund-of-Essex]

Hi I_B,

In 2019/2020 when I choose the transducers (i.e. loudspeakers and digital source) for my HiFi system the Chord DAVE was not included as I wanted as few boxes as possible (i.e. not a separate DAC) and ideally a digital source that might include DSP room correction.

Since then, with my personal experience of the various limitations of Linn products and the domain knowledge I have gained by selecting and applying acoustic room treatments, as well as the shared experience of the Chord DAVE product by Naim forum members, I now consider that 2026 is a good time to revisit the digital source transducer choice I made in 2021.

FYI, the ‘sweetness’ in the treble I heard during my recent experiences in the other locations were in rooms much bigger than my own. So it may have been the room size that helped, or the specific type of room treatment used in both locations (brand was Acustica Applicata), or it might have been the source(s). Chord DAVE was in just one of the two locations…

I’ll keep you posted.

PS: For reference, the comparative ‘bass sound’ in my media room is superior to the ‘bass sound’ I experienced in these other venues - so no help required in that department!