What does Nyquist folding look like?

[wgscott]

I have an HDtracks version of Beethoven's 9th that looks rather good in many ways (summarized here).

 

It looks like the frequency spectrum exceeds 48 kHz, so 96 kHz sampling would produce a folding artifact, but presumably well above the audible threshold.

 

The cross-section reveals some amplitude increasing at the highest sampled frequencies. I was wondering if this is what folding looks like?

 

 

If not, anyone know what it might be? Since it is well above the audible range, this is just a curiosity question...

 

 

 

[Julf]

Well, yes and no. It is not per se how Nyquist folding looks like, but you are likely to get artefacts caused by Nyquist folding if you have signals above half the sample rate. The fact that they are beyond hearing range doesn't help - Nyquist folding is exactly that, folding. It causes artefacts at lower frequencies, by "folding back" the signals both up and down the spectrum. Think of it as beating between the sampling frequency and the signal. In effect, you get the same effect as when a rotating wheel looks like it is rotating much more slowly when lit by a stroboscope, a light driven by an AC frequency, or when recorded by a film camera. Even if the wheel rotates at high speed, you see a much lower rotation speed as a result of the intermodulation of the two frequencies.

 

[CG]

Welcome to my world...

 

Not exactly the same, but this will give you some insight if you fool around with it.

 

http://designtools.analog.com/dt/dac/ad9777image.html

 

Sigma delta converters usually found in audio ADCs and DACs are more complicated because of all the internal calculations done within.

 

[esldude]

Also looking at some HD tracks I get a similar graph. Actually the tracks I have are pretty dead between 20khz and 35 khz. I see the same slightly rising noise that you have here above 35 khz. Even then it is over a 100 dB down from peak. I don't think it is of any concern.

 

Also, if you will change the 'size' parameter to something larger it will slice things up into smaller segments. This will likely drop those peaks on the right hand side a few more dB. Try at least 4096 or 8192 or even 16384.

 

[Miska]

No, that doesn't look like image spectrum.

 

It is usually easy to recognize, since the spectrum is mirrored across Nyquist frequency.

 

In unfiltered case a frequency that is on either side of the Nyquist appears at equal level on the other side, at the same distance in terms of frequency from the Nyquist frequency.

 

Leakage to the wrong side is removed by means of filtering. So if there's any filtering in play, the wrong side should be lower in level.

 

 

[Miska]

Those peaks could be actually coming from an SMPS?

 

 

[barrows]

Maybe... but then they should be present in any analysis dome with the same gear, right?

 

[Miska]

Maybe... but then they should be present in any analysis dome with the same gear, right?

 

Since it's about analysis of a file, yes it should be present in all recordings made with exactly same gear combination and setup.

 

I don't see how the gear used to do the analysis would matter unless it was done from analog domain.

 

 

[Julf]

> Also looking at some HD tracks I get a similar graph.

 

Indeed.

 

> Actually the tracks I have are pretty dead between 20khz and 35 khz.

> I see the same slightly rising noise that you have here above 35 khz.

> Even then it is over a 100 dB down from peak. I don't think it is of any concern.

 

Well, the hf noise in itself is probably not audible. But they might be a sign of something more worrying, that could have effects lower down in the spectrum.

 

To me - and I am happy to be shown the error in my thinking - it looks like the original "master" came from the analog output of a digital (SACD) player with bad HF filtering - that would not be an issue if it was driving an an analog chain, but it then gets redigitized/resampled by whatever recording equipment produced the HDTracks recording, and here it is important to make sure you have no signal content above the Nyqvist frequency, otherwise you run the risk of artifacts and intermodulation noise.

 

The whole benefit of 96 kHz sample rate is that it allows for much softer, gentler filtering that the horrible brickwall filtering needed by 44 kHz - but it does not mean you don't have to do any filtering at all!

 

I have to say that when I found the first hi-res downloads, I (naively) expected them to be direct digital copies of the (digital) master recording (as most modern recordings are recorded, mixed and processed entirely in the digital domain), or, when the original master was analog, a direct digital transfer of the analog tape.

 

What I didn't expect was, in worst case, a recording that was first taken from tape, recorded digitally to DSD fro SACD, then played through a SACD player and it's DAC converter, then (badly) re-digitized at another sample rate (before ultimately being converted one more time to analog through my DAC for listening.

 

[Julf]

Miska,

 

The spectrum mirroring is easy to identify if the signal mirrored is a single tone, but if you have wide-spectrum noise, you would only expect to see a mirrored broad noise spectrum, right?

 

[Julf]

Could be SMPS, but that would be coming from either the original recording equipment or be a sign of a conversion through analog domain later in the chain. I would have assumed that anything in the original recording process would have been cleaned out in the original processing, so SMPS noise would indicate a D/A-A/D resampling/re-digitizing process.

 

[Miska]

The spectrum mirroring is easy to identify if the signal mirrored is a single tone, but if you have wide-spectrum noise, you would only expect to see a mirrored broad noise spectrum, right?

 

..or multiple tones (not noise).

 

Plain noise or noise-like signal is more difficult, but in some cases can be still revealed through detailed enough spectrogram, since it loses some of the randomness that should be there. The Nyquist frequency tends to cause an "edge" there.

 

 

[Miska]

I would have assumed that anything in the original recording process would have been cleaned out in the original processing

 

I think it is in the original recording. And it is as such a low level and high frequency, that it may have been left there. Removing it would have required additional notch filtering pass.