Good morning,
I would like to make a noise measurement using a soundcard with adequate sampling at a good spectral resolution.
Speaking with a radio amateur I was told that according to the Nyquist theorem if e.g. I want to make a measurement on a 50KHz bandwidth, I need at least 96KHz set, i.e. double the sampling frequency, to have a reliable result.
Now maybe mine is a stupid question from a new expert in signal theory, but if the BW of my amplifier on which I carry out measurements was 70KHz, how should I behave in this regard. Should I filter further down to 50KHz or below?
This is because for the calculation of nV/√Hz of interest to me, the frequency response is involved (with a correction factor perhaps if there is no "brick wall").
However, having fiddled around a bit, I have already seen for myself that common soundcards above 20-24KHz then tend to go down a step, due to an LP filter of the DAC, and so on until the end of the bandwidth.
So I wonder if it still makes sense to sample above 20KHz (even if feasible) if the measurement essentially doesn't change much.
Thank you
I would like to make a noise measurement using a soundcard with adequate sampling at a good spectral resolution.
Speaking with a radio amateur I was told that according to the Nyquist theorem if e.g. I want to make a measurement on a 50KHz bandwidth, I need at least 96KHz set, i.e. double the sampling frequency, to have a reliable result.
Now maybe mine is a stupid question from a new expert in signal theory, but if the BW of my amplifier on which I carry out measurements was 70KHz, how should I behave in this regard. Should I filter further down to 50KHz or below?
This is because for the calculation of nV/√Hz of interest to me, the frequency response is involved (with a correction factor perhaps if there is no "brick wall").
However, having fiddled around a bit, I have already seen for myself that common soundcards above 20-24KHz then tend to go down a step, due to an LP filter of the DAC, and so on until the end of the bandwidth.
So I wonder if it still makes sense to sample above 20KHz (even if feasible) if the measurement essentially doesn't change much.
Thank you
I am not good in noise figures, so I cannot comment on that. But I can say something about Nyquist.
If you have a sampling system the sample frequency must be higher than 2 times the highest frequency which is present in the signal.
Note that I emphasize on present in the signal. It is not two time the highest frequency of interest. There must be no detectable frequency components present above 0.5 x sample frequency or Fs. Any signal contents above 0.5 Fs you would find back as a frequency which is mirrored around 0.5 Fs in the frequency spectrum. A frequency of say 0.6 Fs would show up as a frequency of 0.4 Fs in the spectrum. And would be indistinguishable from a real signal in the band.
Source: https://en.wikipedia.org/wiki/Nyquist_frequency
I assume you are using 16-bit A/D conversion. This has a theoretical noise level of 102 dB. So any unwanted components must be 100 dB or so down.
If you are sampling at 96 kHz, theoretically you would be able to sample signals up to 48 kHz. But then you'd need a "brick wall" filter. That is 0 attentuation at 47.9 kHz and infinite (or a 100 dB) at 48.1 kHz. That is impossible.
Specifications are somewhat relaxed if your band of interest is 0-20 Khz. Then any frequencies higher than 48 kHz still would fold back. But as long as the folded back frequencies are higher than 20 Khz you might ignore them. Just make sure that your filter is steep enough that folded back frequencies which end up below 20 kHz are more than 100 dB attenuated.
You'll understand that using a sampling rate of 48 kHz or even 44.1 kHz would impose very high demands on the anti-aliasing filter if your band of interest is 0-20 kHz.
If you have a sampling system the sample frequency must be higher than 2 times the highest frequency which is present in the signal.
Note that I emphasize on present in the signal. It is not two time the highest frequency of interest. There must be no detectable frequency components present above 0.5 x sample frequency or Fs. Any signal contents above 0.5 Fs you would find back as a frequency which is mirrored around 0.5 Fs in the frequency spectrum. A frequency of say 0.6 Fs would show up as a frequency of 0.4 Fs in the spectrum. And would be indistinguishable from a real signal in the band.
Source: https://en.wikipedia.org/wiki/Nyquist_frequency
I assume you are using 16-bit A/D conversion. This has a theoretical noise level of 102 dB. So any unwanted components must be 100 dB or so down.
If you are sampling at 96 kHz, theoretically you would be able to sample signals up to 48 kHz. But then you'd need a "brick wall" filter. That is 0 attentuation at 47.9 kHz and infinite (or a 100 dB) at 48.1 kHz. That is impossible.
Specifications are somewhat relaxed if your band of interest is 0-20 Khz. Then any frequencies higher than 48 kHz still would fold back. But as long as the folded back frequencies are higher than 20 Khz you might ignore them. Just make sure that your filter is steep enough that folded back frequencies which end up below 20 kHz are more than 100 dB attenuated.
You'll understand that using a sampling rate of 48 kHz or even 44.1 kHz would impose very high demands on the anti-aliasing filter if your band of interest is 0-20 kHz.
Normally the sound card should have a built-in anti-aliasing filter. As long as the output noise spectral density of the amplifier does not increase excessively above 20 kHz, a moderate amount of anti-aliasing filtering should suffice. That is, I wouldn't worry about aliasing.
Ok, the explanation is quite clear...so the Alias or mirrored image of the signal produced for F > 1/2*F(Nyquist) (which appears as a folding-back on the band to be examined) can therefore be suppressed with a suitable anti-aliasing filter.
Anyway I should choose a -3dB point low enough for the curve slope to go down, let's say, 20/30dB before the F(Nyquist).
For example, if I wanted to design a filter and I'm interested on a frequency of 1KHz and my F(Nyquist) is 24KHz (for e.g. 48KHz/16bit sampling) I would need I could use a Sallen-Key post-filter:
Anyway I should choose a -3dB point low enough for the curve slope to go down, let's say, 20/30dB before the F(Nyquist).
For example, if I wanted to design a filter and I'm interested on a frequency of 1KHz and my F(Nyquist) is 24KHz (for e.g. 48KHz/16bit sampling) I would need I could use a Sallen-Key post-filter:
Keep in mind what @MarcelvdG wrote in his post. Audio cards are provided with a suitable anti-aliasing filter so you don't have to worry about aliasing. Which is true. You don't have to worry about anti-aliasing filters when using your sound card. I mean, most people are not even aware of a filter in their sound card.
In fact, audio cards use different signal processing techniques to allow the use of a relaxed low-pass filter. The actual sampling rate is much higher than what is presented to you. By using signal processing the high frequencies are filtered out and the signal is presented to you as if it were sampled at 44.1, 48 or 96 kHz.
It still holds that there is a filter. If you are building your own ADC you better be sure to have a filter in place. Although I am sure much more sophisticated ADC's and signal processing libraries are available to build your ADC in the same way as on commercial sound cards.
If you are building your old-fashionded ADC anyway, usually the filters are steeper that what you propose. Having an Fc at 3 Khz to obtain an attenuation of 20 dB at 30 kHz is not that much.
There is a typo in your graph, it states the Fc at 0.3 kHz. You mean 3 kHz.
In fact, audio cards use different signal processing techniques to allow the use of a relaxed low-pass filter. The actual sampling rate is much higher than what is presented to you. By using signal processing the high frequencies are filtered out and the signal is presented to you as if it were sampled at 44.1, 48 or 96 kHz.
It still holds that there is a filter. If you are building your own ADC you better be sure to have a filter in place. Although I am sure much more sophisticated ADC's and signal processing libraries are available to build your ADC in the same way as on commercial sound cards.
If you are building your old-fashionded ADC anyway, usually the filters are steeper that what you propose. Having an Fc at 3 Khz to obtain an attenuation of 20 dB at 30 kHz is not that much.
There is a typo in your graph, it states the Fc at 0.3 kHz. You mean 3 kHz.
Sorry, yes I meant 3 kHz.
I agree that a steeper filter would be better of course but st this point, since the external soundcard already has its own anti-aliasing filter (as you and Marcelvdg say), I don't worry about.
Just for sake of information I can post the hamradio link of the site where I read it from anyway. It could be also that it's me to not have well understood the procedure, and that arising my doubt.
I agree that a steeper filter would be better of course but st this point, since the external soundcard already has its own anti-aliasing filter (as you and Marcelvdg say), I don't worry about.
Just for sake of information I can post the hamradio link of the site where I read it from anyway. It could be also that it's me to not have well understood the procedure, and that arising my doubt.
Yes please do post that article.
If you are using a sound card why would you add your own anti-aliasing filter as well? For this purpose you could consider the sound card as an ideal ADC. As long as you choose your sample frequency sufficiently high the sound card will return a nice spectrum 0-20 kHz. IIRC when you choose the sound card sampling frequency too low (say 8 kHz) you won't get any frequency components higher than 4 kHz. But no aliasing as the sound card takes care of appropriate filtering in the digital domain. To make it appear as it used an adequate and perfect anti-aliasing filter for that sample frequency.
If you are using a sound card why would you add your own anti-aliasing filter as well? For this purpose you could consider the sound card as an ideal ADC. As long as you choose your sample frequency sufficiently high the sound card will return a nice spectrum 0-20 kHz. IIRC when you choose the sound card sampling frequency too low (say 8 kHz) you won't get any frequency components higher than 4 kHz. But no aliasing as the sound card takes care of appropriate filtering in the digital domain. To make it appear as it used an adequate and perfect anti-aliasing filter for that sample frequency.
Here's the link:
https://on4cdu.net/noise-measurements-on-voltage-regulators/
I also asked to the author and he was kind to give some clarification for his setup procedure. In the paragraph Measurements is reported the 50KHz bandwidth used.
https://on4cdu.net/noise-measurements-on-voltage-regulators/
I also asked to the author and he was kind to give some clarification for his setup procedure. In the paragraph Measurements is reported the 50KHz bandwidth used.
I have read the article. I don't know this ADC he is using. However it seems to me that if he selects the sample frequency and the sound card produces a spectrum up to 96 kHz. Although the author states that in reality it is usable up to 60 Khz. Which fits in my idea that the sound card does the anti-aliasing for a much higher sample frequency. And then by signal processing in the digital domain the results are presented like the signal was sampled at 192 kHz with an ideal anti-aliasing filter.
This assumption is supported by the measurement in fig 16 in the article. The author reduces the sample rate to 48 kHz and the spectrum runs up to 24 kHz.
Referring to your original post. When the author wanted a spectrum up to 50 kHz he selected 192 kHz as his sample frequency. When he selected 48 kHz, as a result his bandwidth ran up to 20 kHz.
In any way, apparently when using a sound card for AD conversion the anti-aliasing filter is being taken care of. You should not worry about it.
This assumption is supported by the measurement in fig 16 in the article. The author reduces the sample rate to 48 kHz and the spectrum runs up to 24 kHz.
Referring to your original post. When the author wanted a spectrum up to 50 kHz he selected 192 kHz as his sample frequency. When he selected 48 kHz, as a result his bandwidth ran up to 20 kHz.
In any way, apparently when using a sound card for AD conversion the anti-aliasing filter is being taken care of. You should not worry about it.
Yes, I noticed these two different settings used by the author me too.
Finally, though, he sampled at 48KHz because that one was the bandwidth he was interested to.
I report the specs for the device, this is what I found.
Finally, though, he sampled at 48KHz because that one was the bandwidth he was interested to.
I report the specs for the device, this is what I found.
I assume you are interested in the audio frequency range. Then a decent audio card sampling on 48 kHz should do. Dependent on the implementation and the application you use you'll get the spectrum 0-20 kHz.
What I did not realize when I wrote my first answer to your question was that modern audio cards can be considered as ideal. No concerns about aliasing of filtering. It has all been implemented in a way not visible by the user.
What I did not realize when I wrote my first answer to your question was that modern audio cards can be considered as ideal. No concerns about aliasing of filtering. It has all been implemented in a way not visible by the user.
Are you trying to measure the total noise across the full bandwidth or the noise density at a specific frequency? For audio purposes a limited bandwidth is usually used. And weighting to adjust for hearing sensitivty. Both low frequencies and high frequencies are discarded since human hearing is much less sensitive.
If you want noise density thats a different problem. Typically you use a narrow band filter, correct for the filter skirts (to have an effective brick wall filter) and measure the noise in the band, then correct it for a 1 Hz bandwidth. And then to be useful you correct for the system gain to get the input refered effective noise.
What is your application?
If you want noise density thats a different problem. Typically you use a narrow band filter, correct for the filter skirts (to have an effective brick wall filter) and measure the noise in the band, then correct it for a 1 Hz bandwidth. And then to be useful you correct for the system gain to get the input refered effective noise.
What is your application?
No, I'm not interested only to audio frequency but I haven't the proper (expensive) equipments to do things well. I would like to test up to e.g. some hundreds of KHz to study the behaviour of some common components in term of noise (spectral noise density). I'm not trying to investigate in the RF field because I'm not a radio-amateur.
Let's say that I'd be satisfied for now up to 20KHz too. I saw a lot of graphs on the net of people testing resistors, regulators, diodes etc with more or less at-hand equipments and this intrigues me a little.
I'm also thinking about the best and not much complicated LNA circuit to approach the thing. I created a basic one but also around there are schemes with particular FETs like BF862, 2SK107 etc driving low noise opamps (LT1028, AD797, OPA1611) often configured as transimpedance amplifiers and I thought I'd go for that path. On the contrary, the LNAs with bjt (like ZTX851) offer very low noise at the same time they have low input impedance and therefore have more limited uses.
Let's say that I'd be satisfied for now up to 20KHz too. I saw a lot of graphs on the net of people testing resistors, regulators, diodes etc with more or less at-hand equipments and this intrigues me a little.
I'm also thinking about the best and not much complicated LNA circuit to approach the thing. I created a basic one but also around there are schemes with particular FETs like BF862, 2SK107 etc driving low noise opamps (LT1028, AD797, OPA1611) often configured as transimpedance amplifiers and I thought I'd go for that path. On the contrary, the LNAs with bjt (like ZTX851) offer very low noise at the same time they have low input impedance and therefore have more limited uses.
You might like to try a simple design of Scott Wurcer:
I've built it a long time ago with a 2SK170, it has a noise floor of 0.7nV/rtHz, and I've been using it ever since.
I've built it a long time ago with a 2SK170, it has a noise floor of 0.7nV/rtHz, and I've been using it ever since.
Yes, I notice that interesting thread...a clever solution to avoid bulky capacitors on input signal path.
Unfortunely I don't have available that kind of optocoupler and I can't find either in Ebay. Only Mouser and Digikay. In Aliexpress there are some but I fear they are fake.
Unfortunely I don't have available that kind of optocoupler and I can't find either in Ebay. Only Mouser and Digikay. In Aliexpress there are some but I fear they are fake.
This guy made a simplified version (without the VO1263) of the Wurcer's amplifier:
https://www.mvaudiolabs.com/diy/modern-jfet-noise-measurements/
https://www.mvaudiolabs.com/diy/modern-jfet-noise-measurements/
Thank you for the link, it was unknown to me.
You might also like to have a look at this link:
https://on4cdu.net/noise-measurements-on-voltage-regulators/
You might also like to have a look at this link:
https://on4cdu.net/noise-measurements-on-voltage-regulators/
For a really neat LNA consider the Pearl 3 phono pre and eliminate the RIAA compensation components. Pcb and components available from the Diyaudiostore.
Unless you know the bandwith for which noise is measured, the exercise is useless.
It is very helpful to know higher frequency noise for oscillator supplies. For our ears, the area around 6.3kHz is most important
Unless you know the bandwith for which noise is measured, the exercise is useless.
It is very helpful to know higher frequency noise for oscillator supplies. For our ears, the area around 6.3kHz is most important