Chord just released a new upscaler: Quartet

[sleazyjoe]

No. Higher sampling rate captures higher frequencies. Which we can't hear anyway.

That’s not really what a higher sampling rate is about.


The sampling rate is simply how many times per second the analogue waveform is measured when it’s converted into digital. A higher sampling rate means more samples are taken, giving a more detailed mathematical representation of the original waveform.
Yes, by the Nyquist theorem a higher sampling rate also increases the maximum frequency that can be encoded (44.1 kHz → 22.05 kHz, 96 kHz → 48 kHz, etc.), but that’s only one consequence—not the primary engineering reason for using higher sample rates.


Higher sample rates can also:
• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

• Reduce aliasing generated by DSP processes such as EQ, distortion, compression and synthesis (which is why professional recording is often done at 88.2, 96 or even 192 kHz).


Ultimately, the DAC still reconstructs a continuous analogue waveform from the samples. More samples don’t automatically mean audible improvements, but it’s incorrect to say the only benefit is capturing frequencies we can’t hear.

Imagine digitising a circle. If you record only 20 points around its circumference, software can reconstruct the circle, but it has to estimate more between each point. If you record 2,000 points, the reconstruction requires far less interpolation. Audio works on exactly the same principle: the waveform is sampled, then mathematically reconstructed.

 

[RexrothPigeon]

That’s not really what a higher sampling rate is about.


The sampling rate is simply how many times per second the analogue waveform is measured when it’s converted into digital. A higher sampling rate means more samples are taken, giving a more detailed mathematical representation of the original waveform.
Yes, by the Nyquist theorem a higher sampling rate also increases the maximum frequency that can be encoded (44.1 kHz → 22.05 kHz, 96 kHz → 48 kHz, etc.), but that’s only one consequence—not the primary engineering reason for using higher sample rates.


Higher sample rates can also:
• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

• Reduce aliasing generated by DSP processes such as EQ, distortion, compression and synthesis (which is why professional recording is often done at 88.2, 96 or even 192 kHz).


Ultimately, the DAC still reconstructs a continuous analogue waveform from the samples. More samples don’t automatically mean audible improvements, but it’s incorrect to say the only benefit is capturing frequencies we can’t hear.

Imagine digitising a circle. If you record only 20 points around its circumference, software can reconstruct the circle, but it has to estimate more between each point. If you record 2,000 points, the reconstruction requires far less interpolation. Audio works on exactly the same principle: the waveform is sampled, then mathematically reconstructed.

 

[Mart68]

• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

• Reduce aliasing generated by DSP processes such as EQ, distortion, compression and synthesis (which is why professional recording is often done at 88.2, 96 or even 192 kHz).

None of these things are relevant to playback since they're all inaudible with any competent convertor at 44.1KHz. Higher sampling rates do have a use in recording, they're redundant for playback.

Probably you won't agree but the evidence is against you wrt audibility.

 

[dtaylo1066]

Tap dancing.

 

[Purité Audio]

I bet Monty never thought when he made that video…
Keith

 

[LTig]

That’s not really what a higher sampling rate is about.

It is, for playback.

The sampling rate is simply how many times per second the analogue waveform is measured when it’s converted into digital. A higher sampling rate means more samples are taken, giving a more detailed mathematical representation of the original waveform.
Yes, by the Nyquist theorem a higher sampling rate also increases the maximum frequency that can be encoded (44.1 kHz → 22.05 kHz, 96 kHz → 48 kHz, etc.), but that’s only one consequence—not the primary engineering reason for using higher sample rates.

You may not be aware but the sampling theorem states that a band limited signal is reconstructed 100% perfect - this is mathematically proven. If you don't care about inaudible frequencies above 20 kHz the 44.1 kHz sample rate is perfectly sufficient. Higher sample rates will not lead to some imaginable better sound.

Higher sample rates can also:
• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

That's why during recording you use higher sample rates, and at playback DACs upsample internally to higher sample rates.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

This information is required for the inaudible frequencies to be reconstructed, not for audible frequencies.

• Reduce aliasing generated by DSP processes such as EQ, distortion, compression and synthesis (which is why professional recording is often done at 88.2, 96 or even 192 kHz).

That's why recording and mastering is done with higher sample rates. It's not required for playback.

Ultimately, the DAC still reconstructs a continuous analogue waveform from the samples. More samples don’t automatically mean audible improvements, but it’s incorrect to say the only benefit is capturing frequencies we can’t hear.

Imagine digitising a circle. If you record only 20 points around its circumference, software can reconstruct the circle, but it has to estimate more between each point. If you record 2,000 points, the reconstruction requires far less interpolation.

To construct a circle 3 data points are sufficient. Adding more does not make the circle more round.

Audio works on exactly the same principle: the waveform is sampled, then mathematically reconstructed.

There is no interpolation in the traditional sense (like using splines) between data points, not even during upsampling. Here the missing data points are nulled and the digital lowpass filter calculates their final values. After the DAC the analog reconstruction filter removes the steps and thereby creates the analog signal as it was digitized.

Richard Lyons Understanding Digital Signal Processing is a good read, see chapter 10 Sample rate Conversion.

 

[Chester]

To construct a circle 3 data points are sufficient.

Wouldn’t that be a triangle?

I know absolutely nothing about this stuff so happy to be told I’m wrong.

 

[voodooless]

That’s not really what a higher sampling rate is about.


The sampling rate is simply how many times per second the analogue waveform is measured when it’s converted into digital. A higher sampling rate means more samples are taken, giving a more detailed mathematical representation of the original waveform.
Yes, by the Nyquist theorem a higher sampling rate also increases the maximum frequency that can be encoded (44.1 kHz → 22.05 kHz, 96 kHz → 48 kHz, etc.), but that’s only one consequence—not the primary engineering reason for using higher sample rates.


Higher sample rates can also:
• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

• Reduce aliasing generated by DSP processes such as EQ, distortion, compression and synthesis (which is why professional recording is often done at 88.2, 96 or even 192 kHz).


Ultimately, the DAC still reconstructs a continuous analogue waveform from the samples. More samples don’t automatically mean audible improvements, but it’s incorrect to say the only benefit is capturing frequencies we can’t hear.

Imagine digitising a circle. If you record only 20 points around its circumference, software can reconstruct the circle, but it has to estimate more between each point. If you record 2,000 points, the reconstruction requires far less interpolation. Audio works on exactly the same principle: the waveform is sampled, then mathematically reconstructed.

You just failed sampling theorem 101!

 

[LTig]

Wouldn’t that be a triangle?

I know absolutely nothing about this stuff so happy to be told I’m wrong.

There is only one possible circle which will cover those 3 points.

In digital audio there is only one analog curve which covers the sampled points. It is created by the reconstruction filter.

 

[pablolie]

Btw, if you go to Dave page, they are still saying "latest gen FPGA". Yeah, latest gen on a 11yo product

While I very much doubt their FPGA-based filter stuff does anything I'd care about, the advantage of FPGAs is you can keep upgrading their firmware and thereby "upgrade" the functionality, so probably they mean the FPGA + firmware combo?

 

[SSS]

Wouldn’t that be a triangle?

I know absolutely nothing about this stuff so happy to be told I’m wrong.

You are right. Both a triangle and a circle can be constructed by 3 points. In case of the circle the outline of the circle should meet all 3 points. This can be done by de- or increasing the circle diameter until it fits.

 

[pablolie]

You are right. Both a triangle and a circle can be constructed by 3 points. In case of the circle the outline of the circle should meet all 3 points. This can be done by de- or increasing the circle diameter until it fits.

Totally out of topic, but with only 3 reference points you'd have to be very lucky to guess the right triangle shape (unless you're given extra info). With a circle or a perfect square, three random points on the line(s) is all you need (for a square they need to be on different lines). Draw 2 lines connecting the points, measure and draw lines at a right angle from the center of each line... wherever they intersect is the middle of your circle or square. Has little to do with Shannon-Nyquist, though...

 

[SSS]

Totally out of topic, but with only 3 reference points you'd have to be very lucky to guess the right triangle shape (unless you're given extra info). With a circle or a perfect square, three random points on the line(s) is all you need (for a square they need to be on different lines). Draw 2 lines connecting the points, measure and draw lines at a right angle from the center of each line... wherever they intersect is the middle of your circle or square. Has little to do with Shannon-Nyquist, though...

Indeed off topic. But also the whole discussion here about better or even better reconstruction filters is without real benefit at least for me. Where the main sound is below say 8 kHz there are more than 2 data points for reconstruction. Above there is anyway not much melody content. It is sound from cymbals, snare drum etc. which is more like noise. And yes, better noise may feel better.
A sh... storm may blow me away ....

 

[solderdude]

The sampling rate is simply how many times per second the analogue waveform is measured when it’s converted into digital.

Yes.

A higher sampling rate means more samples are taken, giving a more detailed mathematical representation of the original waveform.
Yes, by the Nyquist theorem a higher sampling rate also increases the maximum frequency that can be encoded (44.1 kHz → 22.05 kHz, 96 kHz → 48 kHz, etc.), but that’s only one consequence—not the primary engineering reason for using higher sample rates.

It is the primary reason, extended FR..

Higher sample rates can also:
• Reduce the steepness of the anti-aliasing and reconstruction filters, making them easier to design and potentially reducing unwanted artefacts.

Nope for higher sample rates the same steepness reconstruction filter is required.

What oversampling can be handy for (using a steep digital reconstruction filter) is the post filter can be simpler as the 'steps' are smaller in amplitude and higher in frequency so less steep post filtering is required.

• Improve the accuracy of time-domain reconstruction because the DAC has more information to work with between samples before interpolation.

That's what math is for.

Ultimately, the DAC still reconstructs a continuous analogue waveform from the samples.

Yes.

More samples don’t automatically mean audible improvements, but it’s incorrect to say the only benefit is capturing frequencies we can’t hear.

It is not incorrect, a higher sample rate has the benefit of capturing higher frequencies.

Imagine digitising a circle. If you record only 20 points around its circumference, software can reconstruct the circle, but it has to estimate more between each point. If you record 2,000 points, the reconstruction requires far less interpolation. Audio works on exactly the same principle: the waveform is sampled, then mathematically reconstructed.

There is no 'estimating' involved but math.
Reconstruction is not estimation but calculation.
Therein lies the difference.

 

[Chester]

So, parking for one moment the fact that it’s inaudible, which I think most of us here agree on, is Rob Watt’s overarching approach theoretically correct or incorrect?

If the waveform can be perfectly reconstructed from 44.1khz and a handful of taps, how can he claim his approach is “more perfect“ ?

I guess it’s probably not as black and white as I would like it to be, which is a shame as it would put this whole thing to bed.

 

[Mart68]

So, parking for one moment the fact that it’s inaudible, which I think most of us here agree on, is Rob Watt’s overarching approach theoretically correct or incorrect?

If the waveform can be perfectly reconstructed from 44.1khz and a handful of taps, how can he claim his approach is “more perfect“ ?

I guess it’s probably not as black and white as I would like it to be, which is a shame as it would put this whole thing to bed.

It's not possible to do a perfect reconstruction no matter how many taps, it can only get closer to perfect.

The question is 'what is the audible limit', since going beyond that is pointless.

Watts claims that taking this to extremes is audible and beneficial to sound quality, however he won't publicly blind test to vindicate his claim. Which makes me suspect he has attempted it privately, and failed.

 

[RandomEar]

As far as I know, Chord filters are closing in on four million taps, right? A typical 44.1 kHz recording of three minutes only has roughly 8 million samples. Their filter - as far as I understand - can't be linear phase due to the delay of roughly 45 s it would incur otherwise. Putting aside the irony that this makes the reconstruction inherently less faithful and also means that up to about 45 s into the song the filter won't be fully engaged because it's missing samples: Is it really useful to "look back" that amount of time to reconstruct the current newest sample being filtered? Like, is it plausible that you need to take into account half of the whole f*cking song to correctly reconstruct one single sample?

EDIT: Forgot about the upsampling. So likely, you've got to divide all numbers by something like 16, making it initially less absurd. Linear phase would also be possible in that case. The absurdity of having 8 million actual samples vs 4 million taps still stands.

 

[sq225917]

It'd be interesting to see tap length plotted against deviation from the original analogue signal.

Ie compare cd spec output, vs the same original input file sampled at say 100x the original sample rate. Then see how much the cd spec deviates from the mega sampled original at diffetent tap lengths.

I'd assume its a log curve that hits the square root of sweet f-ck all pretty quickly.

 

[RandomEar]

It'd be interesting to see tap length plotted against deviation from the original analogue signal.

Ie compare cd spec output, vs the same original input file sampled at say 100x the original sample rate. Then see how much the cd spec deviates from the mega sampled original at diffetent tap lengths.

I'd assume its a log curve that hits the square root of sweet f-ck all pretty quickly.

That would look roughly like this:

"Ground truth" is the original synthetic signal consisting of a mix of four sine waves up to 16 kHz. "Downsampled" is what would be pressed onto a CD, ignoring the 16 bit quantization. "LIN fast (N taps)" are the reconstructed signals based on the downsampled datapoints using linear phase fast roll-off filters with 64 up to 512 taps and 8x oversampling. I took the original filter design from here and it is using 512 taps. So all other filters are not using optimized settings like pass band, stop band, ripple and so on. This is the absolute worst case comparison - you could easily get a significantly better filter with 64 taps if you allowed for a slightly broader transition region, a tiny bit more ripple in the pass band or a bit less suppression above Nyquist.

But even so, it's pretty clear that 512 and 256 taps are visually indistinguishable from the ground truth. If you allowed for higher frequency signals of up to 22.05 kHz in the mix, maybe you would see some small differences between the best two filters in some regions. But there's very little of those frequencies in actual music.

 

[danadam]

It'd be interesting to see tap length plotted against deviation from the original analogue signal.

Here are 3 filters with different tap lengths:

Chord quest, Vs Rme adi-2 dac fs + tap count.

Well it looks a lot more advanced than this. But is it really?

www.audiosciencereview.com