I just posted a blog entry "what is going on and why is this?"

Link to blog site post

have a read if you are interested and do the poll please ?

The poll will help me on my future posts 🙂

.

- Home
- General Interest
- Everything Else
- "Noise Floor" and S/N ratio in and with FFT graphs

You are using an out of date browser. It may not display this or other websites correctly.

You should upgrade or use an alternative browser.

You should upgrade or use an alternative browser.

- Thread starter dddac
- Start date

I just posted a blog entry "what is going on and why is this?"

Link to blog site post

have a read if you are interested and do the poll please ?

The poll will help me on my future posts 🙂

.

For a long time I was VERY confused about the FFT baseline, and how it relates to noise. However, the FFT "noise floor" thing becomes quite a bit easier once you remember that each point on the FFT spectrum shows the signal amplitude *in the corresponding frequency bin*. The baseline values therefore show the noise in each frequency bin. More bins means smaller bins, and hence lower baseline values. Total noise is the sum of all baseline values.

Easy 🙂

I downloaded a text file with a macro which will do it in (almost) real time, but I find that appending many runs and RMS'ing the values in Excel gives a more accurate estimate of the noise floor. Armed with this data, you can run the FFT compare and create a factor at each frequency which makes the two data sets congruent.

BTW, the RMS detector of the AP should be 4 samples per second for best noise measurement, and it helps to reduce the bandwidth to 80kHz.

correct, there is a link to this paper in the post.

I also use an Excel File to run it like that. Made some time ago. Excel is predestined for that of course 🙂 It is a pity there is no macro for good old 2322 system two. There is one for the 2722 I believe...

Thanks for confirming the 4 samples / sec. I use that as well, but more from the gut ...

Good point an the analog input. I will run that sweep as you suggest and see what comes back

what is behind 0,4812?

If this is a "large part of the problem", I am not sure, but at least it is now the right wording.

I hope the article itself will be helpful for those not so deeply in the matter.

what is behind 0,4812?

It's a factor which takes into account the "skirts" of the bandpass filter. Most of the time in audio we look for the -3dB points. However, in noise measurement, you are instead trying to take into account all of the "power".

A simple way to get your arms around the concept -- if you have access to a very low noise JFET amplifier, perhaps a pair of LSK170 with a battery supply set for a gain of 10. You can reduce the noise floor to ~800pV/Rt Hertz. If you put a 10k resistor onto the input the amplifier will yield a noise voltage of ~12.7nV/Rt Hertz. Run the bandpass test, download into Excel, divide by (SQRT frequency * 0.4812) and you should get a chart with a nearly flat Y-value of 12.7nV/Rt Hz.

Walt Jung mentioned it in his 1995 article on noise measurement of linear regulators. I confirmed with the folks at AP that the filter factor for WJ's System One, and my 2722 were the same.

A further reference -- Quan-Tech which used to manufacture the 5173 Semiconductor Noise Analyzer used a 10k resistor as "sanity check".

Gerhard Hoffman uses a 60 ohm resistor which yields 1nV/Rt Hz, Scott from Tavish Design uses 390 ohms for 2.5nV/Rt Hz.

The macro for the 2722 should work in the 2322. You will also find that using the AP "Scaling for Noise" FFT macro which adjusts for bin width and filter function -- the results come out almost exactly the same.

Last edited:

I am using the 22A with a modified EMU0404 USB for FFT, BTW - hooked up to the monitor outputs of the 22A.

In the process of designing a board which will use the OnSemi NSVJ5908 dual LN JFETs. These perform a bit worse than the BF862, but they are in current production and only cost $0.75 in quantities of 10. After the boards are tested, will offer them at cost in respect of Groner's IP.

well sorry for my poor English 😛 - thanks for the tip - I fixed it now.

You have indeed fixed it. But I don't think its a matter of 'poor English' as plenty of native English speakers make the same mistake. Its not simply about terminology. The pictures you show (I've started reading now) do indeed look like graphs, not like histograms. You even confirm that - 'the curve values you see....'. Curves are characteristics of graphs, not of histograms.

I therefore suspect merely changing 'graphs' to 'histograms' doesn't get us to a solution, it might be more confusing without a bit more explanation. So I suggest perhaps saying 'these plots do look like graphs but the underlying data they're derived from is not continuous. Its discrete - so any appearance of dots being joined together in curves is highly deceptive.' Or something like that.

Its the same problem that happens when people join up time domain samples in a digital recording either using lines or sometimes staircases to make 'a waveform'.

If this is a "large part of the problem", I am not sure, but at least it is now the right wording.

My apologies for sacrificing clarity on the altar of brevity. My meaning was more along the lines of 'the large part of the problem isn't just in using the correct terms but also understanding the graphs as being in reality histograms in disguise'.

You could also show overlays of bar charts from different FFT lengths to illustrate how the baseline gets lower with smaller bins, and how the sum of of the small bins contained in a large bin corresponds to the value of the large bin (use a linear y axis scale for this, not dB).

I personally would not use the histogram term, as I feel this smells too much like statistical distribution functions; but I can see where the idea comes from.

Speaking about terminology: from the audio testing perspective, there is hardly any difference between FFT and 'discrete Fourier transform' (DFT). The FFT is just a special routine to compute the DFT that runs faster than other DFT routines.

I personally would not use the histogram term, as I feel this smells too much like statistical distribution functions; but I can see where the idea comes from.

Do you and/or Marcel have a suggested replacement word which maintains the discreteness? I'm all ears😛

Depends whether you're talking of the data treatment (data bins, binning), or the graphical display of the data (bar chart, staircase plot). But, as I already wrote, the statistics smell of the "histogram" term may be a personal thing in my brain.

Oh, the "graph" term is wired as something very generic in my brain. It does not specifically pertain to continuous ("analogue") data at all. A graph is just a graphical representation of data, no matter if the data are continuous or discrete. According to my brain, bar charts, staircase plots and histograms are all "graphs".

Oh, the "graph" term is wired as something very generic in my brain. It does not specifically pertain to continuous ("analogue") data at all. A graph is just a graphical representation of data, no matter if the data are continuous or discrete. According to my brain, bar charts, staircase plots and histograms are all "graphs".

Last edited:

We start missing the point of my blog which was to help the typical DIY with some kind of interpretation of the FFT "graphs, histograms or whatevers..." in relation to getting an idea of the S/N ratio which belongs to it

This start to remind me of the 92 possible genders where I know only two from Biology class. I must be the dummy here

ooppss Shitstorm of the LGQRSTXYZBLM coming 🙄

In the process of designing a board which will use the OnSemi NSVJ5908 dual LN JFETs. These perform a bit worse than the BF862, but they are in current production and only cost $0.75 in quantities of 10. After the boards are tested, will offer them at cost in respect of Groner's IP.

YES I am Jack , I am indeed looking for an extra LN, small signal amplifier as measurement amplifier as front end for the AP2322. just make a nice thread on it?

I experimented a bit and what I struggle with is that if you want very high input impedance like >> 1Mohm for precisely measuring filter response in real life it depends soo much on the input capacitance of the FET (plus Miller effect) that you only have a high input impedance up to a few kHz....

The 100k Input impedance of the AP are sometimes a limiting factor

any one with ideas are welcome to share of course

Last edited:

Do you and/or Marcel have a suggested replacement word which maintains the discreteness? I'm all ears😛

Finite set of discrete points?

I experimented a bit and what I struggle with is that if you want very high input impedance like >> 1Mohm for precisely measuring filter response in real life it depends soo much on the input capacitance of the FET (plus Miller effect) that you only have a high input impedance up to a few kHz....

The 100k Input impedance of the AP are sometimes a limiting factor

I ran some impedance tests this afternoon -- Groner did not have a chart in his LA article. See this thread. https://www.diyaudio.com/forums/equ...-audio-vol-3-spare-boards-29.html#post6431001

We start missing the point of my blog which was to help the typical DIY with some kind of interpretation of the FFT "graphs, histograms or whatevers..." in relation to getting an idea of the S/N ratio which belongs to it

I see the point of your blog and agree that its dealing with a valuable issue here. Maybe its my nerdy-teacher background coming to the fore - I like to delve into what's behind people's misunderstandings and how they come about rather than simply correct them.

Carry on the good work anyway - I've changed my mind having seen your context - I think it would be less misleading to use 'plot' than 'histogram'.

- Home
- General Interest
- Everything Else
- "Noise Floor" and S/N ratio in and with FFT graphs