4 ms·

So it's been many years since I looked at this stuff, but the reason for higher sample rate isn't to do with humans being able to hear over 20khz. If you sample

by doix 4y ago

So it's been many years since I looked at this stuff, but the reason for higher sample rate isn't to do with humans being able to hear over 20khz. If you sample at 44khz you need a high pass filter at 22khz to avoid aliasing. Since that's pretty close to 20khz, it needs a pretty sharp drop off. Having a filter with a sharp drop off can introduce artifacts. With 96khz sampling rate, you don't need as a sharp drop off in your filter. The above is mostly accurate, hopefully someone else will comment and tell me how I'm wrong and correct the inaccuracies. Edit: to anyone reading this, read the replies if you want the correct explanation. To everyone that replied, thank you :).

kazinator 4y ago
You are correct; with a 96 kHz sampling rate, you don't need such a brick wall filter. But then you dohn't have to actually keep the data at 96 kHz; it can be reduced to 44, this time using a digital filter. A sharply cutting off analog filter is expensive to produce. It has multiple stages to create the multiple poles. High precision resistors and capacitors have to be used to get all those circuit stages to line up. The filter will have phase distortion. That's the basis of "supersampling": sampling at a higher rate with a simpler filter with less of a cutoff, then completing the job with a digital filter to get to the target sample rate. This can be done in reverse, in reproducion. Take, say, a 48 KHz signal, and digitally interpolate it to a higher sample rate like 96 KHz. That is fed to the DAC. The filter after the DAC then doesn't need such a steep cutoff after 20 KHz. A greater bit depth can be used; like 16 bit samples interpolated to 24 bit at a higher rate, fed to a 24 bit DAC. The digital filter or interpolator doesn't care about accurate resistors, capacitors or drift in component values over time or due to heat; it does the same thing with the same data every time. Digital filters can look at future values also. The state of an analog filter is determined by only the current and past values of the signal; but digital signal processing can delay the signal a little bit and look at a "box" around the current value. I think that is key to preserving phase relationships.
- PaulDavisThe1st 4y ago
  > with a 96 kHz sampling rate, you don't need such a brick wall filter. This is incorrect. See my comment above/below.
TylerE 4y ago
Yes, i hate when people talk about Nyquist as if its some Panacea. Its vwry “spherical cow”. It assumes all your filters, DACs, are perfect, and that time is infinite.
- masklinn 4y ago
  The exact same spherical cow issue applies, with worse effects, when you increase your sampling rate unnecessarily. And you can use oversampling to correct for the filtering issues. That's pretty standard in modern DACs and ADCs.
- PaulDavisThe1st 4y ago
  Nyquist isn't about such assumptions at all. It is an information theoretic value, based on the provable claim that there is zero information present in the original analog waveform that is not represented in the digital version. Yes, your filters, DAC and speaker wire could still create a less than ideal listening situation, but the fundamental aspect of the Nyquist frequency is not concerned with any of that.
  - kazinator 4y ago
    There is zero information present in a waveform that is not contained in the digital version, if that original waveform is confined below the Nyquist limit. Either it is that way already, or else is derived from an original-original waveform that isn't, by low-pass filtering.
    - PaulDavisThe1st 4y ago
      That's an excellent coda/qualification of what I said. Put into a more listener centric context: there's no audible information in a waveform that is not captured by sampling it at 44.1khz.
- Dylan16807 4y ago
  You only need so much margin, though.
masklinn 4y ago
Anything above 44 is problematic because lots of systems have non-linear responses and will create audible intermodulations of the ultraharmonics. Monty (of xiph) has an article on the subject (well on 24/192 but it applies all the same) which goes into the gory details: https://people.xiph.org/~xiphmont/demo/neil-young.html https://people.xiph.org/~xiphmont/demo/neil-young.html
- pnut 4y ago
  Interesting video on sample rate effects during digital music production https://youtu.be/-jCwIsT0X8M https://youtu.be/-jCwIsT0X8M
ajross 4y ago
That's the "further processing" notion the grandparent comment talked about. A correctly processed signal at 44.1kHz is more than sufficient to reproduce all audio. But if you're going to do something with the sample series (like low-pass it for aliasing protection in your example, or to resample it for conversion, or for any effects or mixing step really) it's good to have extra data in there to prevent headaches. Likewise sampling at 24+ bits prevents accumulated error in repeated processing, etc... There's a space in the audio world for more bits of data. But the final output format isn't where it belongs.
NovemberWhiskey 4y ago
That's fine - but, to the extent that's true, it's true only about the recording and mastering process. CD, as a delivery mechanism, is totally fine and perceptually flawless.
PaulDavisThe1st 4y ago
[ EDIT: contents of this post were factually incorrect. What Kazinator wrote in response is correct, and I don't want to leave incorrect ideas floating around. ]
- mturmon 4y ago
  This explanation loses me at "infinity per dB" ... and I'm an EE. I think you're trying to cover too much ground here, it's really confusing to try to understand what you mean. I believe the comment you're responding to is talking about the analog filter that is needed to avoid aliasing -- as the first words of your comment correctly note/explain. And in particular, the original comment seems to be noting the phase distortion (in frequencies near the cutoff) that analog brick-wall filters will cause. This has been a design contention for decades, really, ever since the CD format was introduced. It's a big design space, with options for gentler analog filters, followed by very fast digital sampling, and further tricks with filtering in the digital domain, where you don't have to worry about getting great capacitors, etc. It may be out-of-scope to lay all that out in one paragraph!
  - PaulDavisThe1st 4y ago
    That's why I had some blank lines in there :) You're right that its a big design space. The key takeaway is that "yes, higher sample rates can actually make a difference, but almost entirely down to the filter design, not because Nyquist moves ... and you probably cannot hear the difference."
- kazinator 4y ago
  If the goal is audio capture and reproduction, the filter used when sampling at 96 kHz still starts rolling off at around 20 kHz. The wrap-around aliasing artifacts do not begin until around half of 96 kHz, or around 48 kHz. So the filter only has to be steep enough to hit a large attenuation at 48 kHz: full signal to virtually nothing over the space of 28 kHz. If you're sampling at 48 kHz, the signal has to be severely attenuated already; it has to go from 20 kHz to deep cutoff in just the space of a few kHz. At 96 kHz can achieve the effect as if you were sampling at 48 kHz, with a steep filter. You sample at 96 kHz with a milder filter, and then purely in the digital realm, you down-sample to 48 kHz. There is an overall filter consisting of the original analog one plus the digital processing. That is cheaper and more reliable than doing it all in analog. An analog filter with a steep cut off will be challenging in mass production because of the strict component tolerances. Sure, you could use a steep filter with 96 kHz also. Say, a steep filter that starts cutting off at 30 kHz. It would still be a less demanding filtering application because of the margin that you have in the frequency domain. The multiple poles of the filter don't have to be lined up as well. E.g. if the first pole starts rolling off at around 30 kHz, and then next ones at 31, and the third one at 28, ... it doesn't matter because you're still hitting the absolute target of there being next to nothing at 48 kHz, and nearly the full signal at 20 kHz.

by doix 4y ago

kazinator 4y ago

You are correct; with a 96 kHz sampling rate, you don't need such a brick wall filter. But then you dohn't have to actually keep the data at 96 kHz; it can be reduced to 44, this time using a digital filter. A sharply cutting off analog filter is expensive to produce. It has multiple stages to create the multiple poles. High precision resistors and capacitors have to be used to get all those circuit stages to line up. The filter will have phase distortion. That's the basis of "supersampling": sampling at a higher rate with a simpler filter with less of a cutoff, then completing the job with a digital filter to get to the target sample rate. This can be done in reverse, in reproducion. Take, say, a 48 KHz signal, and digitally interpolate it to a higher sample rate like 96 KHz. That is fed to the DAC. The filter after the DAC then doesn't need such a steep cutoff after 20 KHz. A greater bit depth can be used; like 16 bit samples interpolated to 24 bit at a higher rate, fed to a 24 bit DAC. The digital filter or interpolator doesn't care about accurate resistors, capacitors or drift in component values over time or due to heat; it does the same thing with the same data every time. Digital filters can look at future values also. The state of an analog filter is determined by only the current and past values of the signal; but digital signal processing can delay the signal a little bit and look at a "box" around the current value. I think that is key to preserving phase relationships.

PaulDavisThe1st 4y ago
> with a 96 kHz sampling rate, you don't need such a brick wall filter. This is incorrect. See my comment above/below.

TylerE 4y ago

Yes, i hate when people talk about Nyquist as if its some Panacea. Its vwry “spherical cow”. It assumes all your filters, DACs, are perfect, and that time is infinite.

masklinn 4y ago
The exact same spherical cow issue applies, with worse effects, when you increase your sampling rate unnecessarily. And you can use oversampling to correct for the filtering issues. That's pretty standard in modern DACs and ADCs.
PaulDavisThe1st 4y ago
Nyquist isn't about such assumptions at all. It is an information theoretic value, based on the provable claim that there is zero information present in the original analog waveform that is not represented in the digital version. Yes, your filters, DAC and speaker wire could still create a less than ideal listening situation, but the fundamental aspect of the Nyquist frequency is not concerned with any of that.
- kazinator 4y ago
  There is zero information present in a waveform that is not contained in the digital version, if that original waveform is confined below the Nyquist limit. Either it is that way already, or else is derived from an original-original waveform that isn't, by low-pass filtering.
  - PaulDavisThe1st 4y ago
    That's an excellent coda/qualification of what I said. Put into a more listener centric context: there's no audible information in a waveform that is not captured by sampling it at 44.1khz.
Dylan16807 4y ago
You only need so much margin, though.

masklinn 4y ago

Anything above 44 is problematic because lots of systems have non-linear responses and will create audible intermodulations of the ultraharmonics. Monty (of xiph) has an article on the subject (well on 24/192 but it applies all the same) which goes into the gory details: https://people.xiph.org/~xiphmont/demo/neil-young.html https://people.xiph.org/~xiphmont/demo/neil-young.html

pnut 4y ago
Interesting video on sample rate effects during digital music production https://youtu.be/-jCwIsT0X8M https://youtu.be/-jCwIsT0X8M

ajross 4y ago

That's the "further processing" notion the grandparent comment talked about. A correctly processed signal at 44.1kHz is more than sufficient to reproduce all audio. But if you're going to do something with the sample series (like low-pass it for aliasing protection in your example, or to resample it for conversion, or for any effects or mixing step really) it's good to have extra data in there to prevent headaches. Likewise sampling at 24+ bits prevents accumulated error in repeated processing, etc... There's a space in the audio world for more bits of data. But the final output format isn't where it belongs.

NovemberWhiskey 4y ago

That's fine - but, to the extent that's true, it's true only about the recording and mastering process. CD, as a delivery mechanism, is totally fine and perceptually flawless.

PaulDavisThe1st 4y ago

[ EDIT: contents of this post were factually incorrect. What Kazinator wrote in response is correct, and I don't want to leave incorrect ideas floating around. ]

mturmon 4y ago
This explanation loses me at "infinity per dB" ... and I'm an EE. I think you're trying to cover too much ground here, it's really confusing to try to understand what you mean. I believe the comment you're responding to is talking about the analog filter that is needed to avoid aliasing -- as the first words of your comment correctly note/explain. And in particular, the original comment seems to be noting the phase distortion (in frequencies near the cutoff) that analog brick-wall filters will cause. This has been a design contention for decades, really, ever since the CD format was introduced. It's a big design space, with options for gentler analog filters, followed by very fast digital sampling, and further tricks with filtering in the digital domain, where you don't have to worry about getting great capacitors, etc. It may be out-of-scope to lay all that out in one paragraph!
- PaulDavisThe1st 4y ago
  That's why I had some blank lines in there :) You're right that its a big design space. The key takeaway is that "yes, higher sample rates can actually make a difference, but almost entirely down to the filter design, not because Nyquist moves ... and you probably cannot hear the difference."
kazinator 4y ago
If the goal is audio capture and reproduction, the filter used when sampling at 96 kHz still starts rolling off at around 20 kHz. The wrap-around aliasing artifacts do not begin until around half of 96 kHz, or around 48 kHz. So the filter only has to be steep enough to hit a large attenuation at 48 kHz: full signal to virtually nothing over the space of 28 kHz. If you're sampling at 48 kHz, the signal has to be severely attenuated already; it has to go from 20 kHz to deep cutoff in just the space of a few kHz. At 96 kHz can achieve the effect as if you were sampling at 48 kHz, with a steep filter. You sample at 96 kHz with a milder filter, and then purely in the digital realm, you down-sample to 48 kHz. There is an overall filter consisting of the original analog one plus the digital processing. That is cheaper and more reliable than doing it all in analog. An analog filter with a steep cut off will be challenging in mass production because of the strict component tolerances. Sure, you could use a steep filter with 96 kHz also. Say, a steep filter that starts cutting off at 30 kHz. It would still be a less demanding filtering application because of the margin that you have in the frequency domain. The multiple poles of the filter don't have to be lined up as well. E.g. if the first pole starts rolling off at around 30 kHz, and then next ones at 31, and the third one at 28, ... it doesn't matter because you're still hitting the absolute target of there being next to nothing at 48 kHz, and nearly the full signal at 20 kHz.