CHORD M-Scaler Review (Upsampler)

[Geert]

Why not just ask watts to supply the test results that substantiate his claims as passionforsound has requested.

Say Watts provides an impulse response, would it make us any wiser? Would it answer the question if this IR effectively improves transient timing to the extend that it is audible (potentially even changing timbre and the perception of depth)? Or will we go back to a disagreement about the potential audibility of the problem (microseconds ITD, -301dB noise floor modulation) and the cure? Since we don't believe there's a problem the start with, I'm afraid I know where this is going to lead to.

 

[Dogcoop]

Say Watts provides an impulse response, would it make us any wiser? Would it answer the question if this IR effectively improves transient timing to the extend that it is audible (potentially even changing timbre and the perception of depth)? Or will we go back to a disagreement about the potential audibility of the problem (microseconds ITD, -301dB noise floor modulation) and the cure? Since we don't believe there's a problem the start with, I'm afraid I know where this is going to lead to.

Sorry I wasn’t clear. I was hoping that a watts supplied test would satisfy @PassionforSound and his desire for the impulse response measurement to justify his hearing a substantial improvement in sound quality due to the WTA and its handling of transients.

 

[Geert]

Sorry I wasn’t clear.

No problem, my reply was not meant as critique. Also for PassionforSound the question is what he's going to conclude from the IR. To start with it would already be interesting to see IR's that illustrates an audible timing issue when a high res file is downsampled to 44.1kHz. Anyone can create such a file. If such IR's are not conclusive, then what is Watts IR going to tell us? (Open questions, invite to the real signal processing experts to chime in).

 

[Shadders]

Hi,
Haven't we had this all before with regards to timing and transients reconstruction/processing using special sauce as per MQA ?

All Chord have to do is for a specific dataset such as an excerpt of a song, provide the digital output of the mscaler at 16x, and do the same for a simple zero padding 8x oversampling with reconstruction filter using 100 taps (example), and show the differences between the two.

As long as the data is provided for input, mscaler 16x output, and 8x oversampling output, we can verify the veracity of the mscaler, without Chord giving away any IP. A simple script for matlab or Octave can be used to complete the comparison by third parties.

Anyway, the 16x oversampling is operating on CD quality 16bit with dither, so oversampling is never going to provide stupendous results apart from reducing the noise power by 16x. Not sure why the -300dB is valid of the inherent CD noise despite oversampling processing gain reducing the noise power which is still to 100dB+ more than the -300dB noise floor.

Oversampling - Wikipedia

en.wikipedia.org

Regards,
Shadders.

 

[tmtomh]

This is so clearly wrong I don't know even where to start.

For ages now, all DACs (unless exotic NOS) do internal digital upsampling and reconstruction filtering, creating the required intermediate sample's values.

Why in digital? Because it is almost impossible to make a halfway decent analog filter, notably for low sample rates.
After upsampling and filtering, the analog filter only has to take care for stuff above the upsampled conversion rate.

Thanks for that information. But you didn’t quote my entire comment. You left out the final part, where I made precisely the point that oversampling enabled/eases the implementation of reconstruction filters.

I was attempting to separate the two elements (1. Alleged “recovery of lost information/transients” from massive oversampling; and 2. Benefits of oversampling for reconstruction filter design) for purposes of clarity. If that analytical separation created an incorrect impression, I’m happy for any of those other, more knowledgeable folks I referenced (including you) to provide a better or more holistic explanation.

I do, however, think the main point stands.

 

[Nathan Raymond]

Does Kai Krause work for Chord? Every time I see a Chord product, I think either he must be behind their physical designs or someone at Chord is seriously enamored with his user computer interface design work from back in the '90s:

https://mprove.de/script/99/kai/index.html

 

[Somafunk]

“Does Kai Krause work for Chord?”

His work and GUI reminds me far more of that utterly gawd awful 6moons website and interface

 

[pkane]

That's wrong, we can maximise all of those at a time (the closer we get to true sinc() the better all three properties will become) until calculation artifacts kick in...

For any given filter size you can maximize only some of the parameters. If you keep increasing filter size, you'll very quickly get to the point where computation artifacts limit the precision. Then you'll need to keep increasing filter coefficient precision from 64-bit float to 128-bit float to 256-bit float, and beyond as you keep increasing the number of taps.

Not sure what exactly this improves but if you do it in real time, you'll also need to wait for a very long time for the music to start playing. And of course, you'll never get to truly "perfect" reconstruction, not until maybe quantum computing becomes a reality

 

[bennetng]

What constitutes best, though? Is it stopband attenuation? Passband flatness? Transition sharpness? Since we can't maximise all of those at the same time, what compromises are acceptable? Rob Watts takes a very simplistic "more taps better" view, and like the sheep in Animal Farm, his acolytes are happily bleating this mantra. It is, however, anything but clear that his approach has any merits compared to traditional methods.

Rob Watts' "window" is not too different from rectangular. A very rectangular looking Tukey?

Chord DAVE Measurements (With MScaler) - GoldenSound

Technical measurements and analysis of the Chord DAVE DAC. This is a high end DAC with some great and not so great aspects

goldensound.audio

Chord Hugo 2 Review (Portable DAC & HP Amp)

This was my experience. I posted some actual acquired data showing that a headphone accessory made by one of their sponsors did nothing, then took it apart, analyzed the "magic" materials using XPS and SEM, and included the results of that analysis (narrator: They weren't magic). Poof, gone...

www.audiosciencereview.com

 

[Thomas Savage]

Yes, but show me the evidence that you are using to say that it doesn't do anything? It seems that both sides are speculating at this point and it's pointless. If no-one has any actual evidence that it's claimed objective benefits are/are not real then everyone should just go back to enjoying the music and agree that some things can't be measured.

These are high-school level debating tactics you are using . They equate to cheap Jedi mind tricks , FYI these are not weak mined folks .

It is clear you are enamoured by Rob and are doing everything to stop a competing narrative develop within and externally you project these internal workings here in your postings , cognitive dissidence . You've absolutely no chance of being objective as your postings clearly show .

Quite human and in a spirit of generosity I for one will refrain from accusing you of being intellectually dishonest. Others might not be so .. , optimistic.

Enjoy your Chord purchase ( hope you got a good deal ) .

 

[danadam]

The reason is that upsampling simply, and literally, just duplicates samples. So if you upsample from 44.1kHz to 176.4kHz, that's 4x oversampling, which means every original sample is replicated three more times - it's copy-pasted three times so that every original sample is now four samples.

As others have said, that's wrong. I'll try to illustrate (and I'm sure local experts will point out if some of that is wrong ). We start with 1 kHz tone at 8 kHz sampling rate and 8 bits (all generated files are in the attachment):

sox -r8k -c1 -n -b8 "01_sin.8_8.wav" synth 5 sin 1k norm -3

The upsampling is done by inserting zero samples between the existing samples, then filtering the images created by this operation and finally applying a gain to compensate. If we only insert zero samples, we get images in the spectrum:

sox "01_sin.8_8.wav" -r16k "02_sin.8_16.upsample.wav" upsample 2

When we filter out the images and apply the gain, then the zero samples magically mathematically pop up into the expected places:

sox "02_sin.8_16.upsample.wav" "03_sin.8_16.lpf.wav" sinc -t 400 -3800 vol 2

---

What you described is zero-order hold interpolation. It could work, I think, but it is more complicated then what was described above. To show the problem, we start with white noise with 3 kHz bandwidth, again at 8 kHz sampling rate and 8 bits:

sox -r8k -c1 -n -b8 "04_noise.8_8.wav" synth 30 white norm -9 sinc -t 400 -2900

If we only insert zero samples, we get images (same as earlier):

sox "04_noise.8_8.wav" -r16k "05_noise.8_16.upsample.wav" upsample 2

If we filter out the images and apply gain, we'll get the original signal. But if we do zero-order hold instead of zero samples then this happens:

No sox command this time :( If it is possible to do in sox, then I don't know how

We could filter-out the images, but then we are left with this sloping spectrum when the original was flat. From what I understand, we could apply some filter to counteract this effect. So in the end we have this filter vs a simple gain in case of zero samples.

 

[dc655321]

We could filter-out the images, but then we are left with this sloping spectrum when the original was flat. From what I understand, we could apply some filter to counteract this effect. So in the end we have this filter vs a simple gain in case of zero samples.

Yes, there should be a sinc (squared) envelope governing the images. So, invert that to get the “nulling” counter effect filter.

Or just use the zero-stuffing technique for integer up-sampling needs, as you’ve so nicely illustrated

Bonus points for illustrating the (less interesting) integer down-sampling case?!?

 

[tmtomh]

As others have said, that's wrong. I'll try to illustrate (and I'm sure local experts will point out if some of that is wrong ). We start with 1 kHz tone at 8 kHz sampling rate and 8 bits (all generated files are in the attachment):

sox -r8k -c1 -n -b8 "01_sin.8_8.wav" synth 5 sin 1k norm -3

View attachment 220482

The upsampling is done by inserting zero samples between the existing samples, then filtering the images created by this operation and finally applying a gain to compensate. If we only insert zero samples, we get images in the spectrum:

sox "01_sin.8_8.wav" -r16k "02_sin.8_16.upsample.wav" upsample 2

View attachment 220484

When we filter out the images and apply the gain, then the zero samples magically mathematically pop up into the expected places:

sox "02_sin.8_16.upsample.wav" "03_sin.8_16.lpf.wav" sinc -t 400 -3800 vol 2

View attachment 220485

---

What you described is zero-order hold interpolation. It could work, I think, but it is more complicated then what was described above. To show the problem, we start with white noise with 3 kHz bandwidth, again at 8 kHz sampling rate and 8 bits:

sox -r8k -c1 -n -b8 "04_noise.8_8.wav" synth 30 white norm -9 sinc -t 400 -2900

View attachment 220491

If we only insert zero samples, we get images (same as earlier):

sox "04_noise.8_8.wav" -r16k "05_noise.8_16.upsample.wav" upsample 2

View attachment 220492

If we filter out the images and apply gain, we'll get the original signal. But if we do zero-order hold instead of zero samples then this happens:

No sox command this time :( If it is possible to do in sox, then I don't know how

View attachment 220493

We could filter-out the images, but then we are left with this sloping spectrum when the original was flat. From what I understand, we could apply some filter to counteract this effect. So in the end we have this filter vs a simple gain in case of zero samples.

Thank you for the detailed and thorough explanation!

 

[sofrep811]

OK, let's move this conversation forward then. Where is the evidence that the M-Scaler is not improving the accuracy of the timing of transients (or rapid fluctuations of the music waveform - whatever we want to call it) as this is the primary claim from Rob Watts about the benefit of the device.

I agree that the noise floor is higher with the 2x and 4x upsampling modes and I am not going to attempt to debate the subjective and objective merits of triangular and gaussian dither because I am not qualified to. I also think that Chord having these modes active in the M-Scaler is a bit troublesome (as I raised with Rob Watts) and, if I were in charge of Chord, I would probably have had simply 16x or no oversampling and marketed the device entirely for use with Chord DACs.

What I am not seeing is any evidence that the device isn't doing what it claims to do. Before you come back and say that Chord have offered no evidence that it is doing it, that's a different discussion and a reasonable request. My point here is that the M-Scaler is being criticised for not doing anything (or adding extra noise in 2x and 4x modes) when I don't see any measurements actually focussing on what it is claimed to do and other measurements that are objectively irrelevant to its actual performance as confirmed by Amir in his review.

I also don't understand why listening tests weren't conducted at 16x oversampling with the provided Hugo 2 when that is realistically the intended use of the M-Scaler even though it can (but probably shouldn't as I said above) work with other devices.

One final point. I borrowed an M-Scaler ages ago for my review. After spending time listening to it in all modes and using the TT2, I enjoyed its sound (in 16x mode) sufficiently that I purchased it for myself after the review. At the outset of my review process, I had no vested interest in it being good or bad (I have no access to any kind of affiliate programs for any Chord products except sometimes Mojo 2 when it's available via Amazon sellers) but found it to be a very enjoyable addition to the TT2. It's with that experience in mind that I believe it is subjectively improving the listening experience and without evidence to the contrary, I can only assume that it is because of an improvement of the timing information as claimed. There is strong evidence based on studies of auditory processing neurons in the brain that we have very acute timing accuracy in our auditory system so Rob Watts is accurate with his claims about our timing acuity. This IS in my area of expertise. I haven't seen objective data in either direction that proves or disproves this and so I am going with my personal listening experiences at this stage.

Do you have any idea how stupid you sound? After what you’ve wrote in the comments of your own channel and on here? You’re illogical and only waiting around for one bit of data that you can pounce on to negate a solid data set by amir.

What’s your true motivation for defending a device that has no defense unless it was priced strictly for the chassis and parts?

 

[sam_adams]

A forum user's guide to marine mammals.

 

[Marc v E]

Weird and wonderful thread this is. Kind of reminds me of explaining the merits of electric cars by a journalist who has a passion for cars with 'fact-checking' only by an audience that judges by popular vote. Is what the journalist says to be taken as fact? Some of the audience will passionately be against, some for, some don't know and just follow someone else's opinion.

Meanwhile, the engineering facts are that maintenance and driving costs are cut in half. As production numbers go up, the upfront cost of purchasing is getting down. It has less parts and is less complicated to make. Can't be much clearer than that.

To me it's obvious what's going on but to the ones basing their judgements on opinion it's pliable and unsure. Much like the digital waveform and sampling theory. (It's a theory, right?
)I know there's much money to be saved (and gained !) if I can accept the facts and accept I'm wrong if the facts say so, even if I'm dead sure I'm right.

Similarly ,I've been wasting a lot of time in hifi theories and judging by ear alone, forever trying to inprove while running in circles. I'm relieved and glad I found this site through a friend.

 

[DonR]

Weird and wonderful thread this is. Kind of reminds me of explaining the merits of electric cars by a journalist who has a passion for cars with 'fact-checking' only by an audience that judges by popular vote. Is what the journalist says to be taken as fact? Some of the audience will passionately be against, some for, some don't know and just follow someone else's opinion.

Meanwhile, the engineering facts are that maintenance and driving costs are cut in half. As production numbers go up, the upfront cost of purchasing is getting down. I has less parts and is less complicated to make. Can't be much clearer than that.

To me it's obvious what's going on but to the ones basing their judgements on opinion it's pliable and unsure. Much like the digital waveform and sampling theory. (It's a theory, right?
)I know there's much money to be saved (and gained !) if I can accept the facts and accept I'm wrong if the facts say so, even if I'm dead sure I'm right.

Similarly ,I've been wasting a lot of time in hifi theories and judging by ear alone, forever trying to inprove while running in circles. Glad I found this site through a friend.

You need to learn the difference between a scientific theory vs the vernacular use. There is no higher standard in science than a theory (save for mathematical proof).

 

[Marc v E]

You need to learn the difference between a scientific theory vs the vernacular use. There is no higher standard in science than a theory (save for mathematical proof).

I did. Did you miss the winking smiley ? (It's hard to convey irony in text and smileys.)

 

[DonR]

I did. Did you miss the winking smiley ? (It's hard to convey irony in text and smileys.)

My apologies. I am so used to having to point out the difference it has become a reflex reaction. That's my theory, anyhow.

 

[amirm]

Much like the digital waveform and sampling theory. (It's a theory, right?

No. It is a theorem. Note the extra "m." In math this means a proven thing, not a theory. Google for Sampling Theorem and you get this:

As you see, it is mathematical proof ("completely recovered").