It is now quite easy to make measurements of frequency response, distortion and noise using a PC and a good quality soundcard but misinterpretation of the graphs produced has led to some performance claims that sometimes defy the laws of physics especially when it comes to noise.
Here is an example from the "The Wire" thread:
https://www.diyaudio.com/forums/headphone-systems/179298-wire-ultra-performance-headphone-amplifier-pcbs.html
The second post includes an FFT plot from which the author states "the noise floor is down at -145dB" which at first sight appears to be the case as the noise seems to bump along at about -145dB. But the FFT plot is of the noise per root Hz. To get the total rms noise over the entire audio spectrum (which is what you hear and measure with a regular meter) you need to integrate these measurements over the 20KHz bandwidth. This is not complicated to do but if you assume the noise is more or less constant over the entire audio spectrum you can simplify it considerably. All you need do is multiply each measurement by the square root of the bandwidth. The square root of 20,000 is 141 so the rms noise over 20KHz is 141 times the value on the graph or 43dB higher. So the rms noise is -102dB which is still a very good result but not as incredible as the stated -145dB.
Cheers
Ian
Here is an example from the "The Wire" thread:
https://www.diyaudio.com/forums/headphone-systems/179298-wire-ultra-performance-headphone-amplifier-pcbs.html
The second post includes an FFT plot from which the author states "the noise floor is down at -145dB" which at first sight appears to be the case as the noise seems to bump along at about -145dB. But the FFT plot is of the noise per root Hz. To get the total rms noise over the entire audio spectrum (which is what you hear and measure with a regular meter) you need to integrate these measurements over the 20KHz bandwidth. This is not complicated to do but if you assume the noise is more or less constant over the entire audio spectrum you can simplify it considerably. All you need do is multiply each measurement by the square root of the bandwidth. The square root of 20,000 is 141 so the rms noise over 20KHz is 141 times the value on the graph or 43dB higher. So the rms noise is -102dB which is still a very good result but not as incredible as the stated -145dB.
Cheers
Ian
I don't quite understand the bandwidth parameter in RMS noise calculation. We integrate over 20kHz because of the ideal human hearing. But the human perception of frequency is logarithmic. We don't hear zero Hz. Integrating from 0 to 20kHz is not much different from 20Hz-20kHz. And this is not much different from integrating between 100Hz-20kHz, although the subjective noise character is very different.
Also if we integrate between 0 and 10kHz (which is more realistic for persons over 60 like myself), one gets again different noise figure.
Is that noise calculation just a kind of convention, not closely related to real-world noise perception? Is it just for comparison purpose?
Also if we integrate between 0 and 10kHz (which is more realistic for persons over 60 like myself), one gets again different noise figure.
Is that noise calculation just a kind of convention, not closely related to real-world noise perception? Is it just for comparison purpose?
I think it is just the definition of RMS noise: Root of the Mean Squared values.
So you integrate over all the 'Herzes' in the bandwidth, then you take the root.
Perception doesn't come into play here.
Jan
So you integrate over all the 'Herzes' in the bandwidth, then you take the root.
Perception doesn't come into play here.
Jan
Last edited:
Aargh...
We must forget those THD digits as having no correlation with any kind of amplifier (and more - audio itself) sound quality.
We must forget those THD digits as having no correlation with any kind of amplifier (and more - audio itself) sound quality.
The noise shown in the FFT plot is the noise per bin. If the bin's 1Hz wide then yes, it'll be as you've stated. But seeing as the number of points and the sample rate are rather variable then the chances of a bin being 1Hz wide are rather slim. In order to interpret the FFT and determine the effective bin width, the number of points in the FFT, the sample rate and also the windowing function (if any) need to be known.
Oh, well....
This is more than misinterpretation of measurements.
This is plain ignorance.
This is what happens when tools fall in wrong hands, they simply do not know what they are talking about.
This is more than misinterpretation of measurements.
This is plain ignorance.
This is what happens when tools fall in wrong hands, they simply do not know what they are talking about.
I have to echo Markw4's post. While I am sure that this error in noise level was quite annoying for you (Ian) to see, it is in the spirit of this forum to teach others and share ideas, not just bee-atch. A more productive way to express your concern is to start a thread saying something like:
"hey folks, I realized that many people make a mistake when reporting the noise level, and here is a handy way to do it right!"
See how that kind of positive spin and knowledge passing can feel good?
"hey folks, I realized that many people make a mistake when reporting the noise level, and here is a handy way to do it right!"
See how that kind of positive spin and knowledge passing can feel good?
Using true-rms conversion you get an un-weighted noise value. Measuring fhrough a preceding A-law filter you get a corrected rms value that is closer to perception.
The last 14 lines of text, plus Fig. 6, in https://www.analog.com/media/en/training-seminars/tutorials/MT-001.pdf
I couldn't put it any simpler.
I couldn't put it any simpler.
This is not a "human" measurement. It is a meter number. Poorly reported.
Yes, 0-20k, 20-20k, and 200-20k are "all the same" which is why Ian went directly to 141 (strictly 119Hz-20kHz).
I agree completely which is why I said the calculation was complex unless you make some assumptions. However, most current RTAs display FFTs as if the bins were 1Hz wide. However, just about every FFT plotting program is also capable of itself calculating the rms noise from its data taking all this into account. I am pretty sure the AP test set used in the example I gave was capable of this but the author chose not to use it. The point it, this would show the actual noise floor to be considerably higher. than the quoted -145dB
Cheers
Ian
Human perception is another factor entirely. Most current measurement programs will give you the true rms noise over any bandwidth you care to mention. Hoever, because of the square root, the noise in a 10KHz bandwidth is only 3dB lower than in 20KHz.
To account for the response of the ear, various weighting curves can be applied and again most modern programs will do this. The A weighting curve , derived from the Fletcher Munson loudness curves, is commonly used because it tends to improve figures significantly - a key factor for marketing people. Professionals tend to use the C weighting curve which gives a less flattering result. Other version take into account the sensitivity of the ear to peaks in the noise and use peak values rather than rms. All these are covered by various standards.
The point being that unless all these factors are specified, the given figures are somewhat meaningless.
Cheers
Ian
I am sorry my post came over as moaning and whining. That was not my intention which was why I took pains to explain why the example measurement was misinterpreted and to show how, in some case, it can be converted into something more meaningful.
I am more than happy to teach and as others will tell you I do a lot of it on other forums.
Cheers
Ian
I use the term ‘spot noise’ to refer to rt/Hz noise figures and ‘total RMS noise over bandwidth (with the BW specified) ’ to signify the integrated noise.
Indeed, there are some mighty claims made because the two are confounded.
The QA401 does both at the click of a mouse button, so it’s great for doing noise measurements provided you take a few precautions.
Indeed, there are some mighty claims made because the two are confounded.
The QA401 does both at the click of a mouse button, so it’s great for doing noise measurements provided you take a few precautions.
I agree, I'll be more positive.
Forgive me, I am too much in the mood of being mad at snake oil mostly riding on ignorance.
Are you aware that there are currently active forum members who shill for snake oil audio products?
I wish people would stop using 20 Hz to 20 kHz unweighted RMS noise levels altogether. They hardly have any relation to what you hear and optimizing the unweighted 20 Hz to 20 kHz noise often leads to a larger than optimal value for the A-weighted noise or the noise measured according to ITU-R 468. A typical example is the design of phono amplifiers with RIAA correction with a bipolar transistor or a valve input stage. By the way, C weighting is for loud signals as far as I know.
