Suggestions for a many, many-band spectrum analyzer?

Hi collective Prodigy brain,

I'm planning on making a big project in the coming months that would feature a spectrum analyzer, but the finest possible resolution possible, like 64-band, for example, or even higher.

I was at first planning on getting an old graphic EQ and ripping out the crossover circuit, but they only seem to go up to 31-band.

On the other hand, I could just take the incoming audio and split it into 100 channels, with each one having a tuned band-pass filter. This seems like a whole shit-ton of work, pardon my language.

This is a large project, which should (fingers crossed) have a budget of at least a few hundred dollars for the spectrum analyzer. So, really, I'm looking for ease of construction and reliability over cost.

I've heard of people using things like the Arduino to do spectrum analysis, though it's my understanding that it's still fairly primitive. i.e. not capable of 64 or 72-band analysis.

Does anyone have any suggestions or hot leads for me to check out?

thanks much!
Matthew

Yup... Study the Fourier transform, or more specifically fast Fourier transform..

You can divide the frequency spectrum into as many bands as you want.. it just takes more time to sample to get very small frequency buckets, and other pesky details.

They sell DSP chips <$5 but you need to learn to program them..

JR

> old graphic EQ

Analysis filters should be narrower than tone-bending filters.

> graphic EQ ... only seem to go up to 31-band.
> spectrum analyzer,... 64-band, for example, or even higher.

Why stop at 64? With a mere 128 filters you can pick-out each note on the european music scale (12 tones per octave).

The number of L-C filter tanks (or equivalent) goes up as the *square* of the number of bands.

You can do tolerable 1-octave analysis with one L-C per band.

Half-octave analysis needs steeper slopes, and two L-Cs per band.

Third-octave needs 3 L-Cs per band.

You are proposing either very sloppy analysis, or 4,000 L-C filters.

> have a budget of at least a few hundred dollars

I think this will be insufficient.

I know that my octave analyzer, if built today, would run over $100 of parts and mechanicals, for just 8 filter units mux and a pile of LEDs. You are talking many thousands of bucks.

CPU is different but similar. The most direct way is a DSP with filters programmed into it. I suspect the brain-power available will not support a large number of narrow filters per DSP, so you have rows of fairly complicated chips. FFT is another route, but does not directly give octave-like bands, it is a linear process, more detail at 9KHz than at 38Hz. My old FFT software declines to even try heavy FFT in real time, because older CPUs would just burn-up or stall. A Q660 might do a whole lot better, but sixth-octave filtering in the bass is just not likely.

Get a piano soundboard. Shine a laser past each string, project on wall. Tune the strings. The relative blur will tell the relative amplitude at that frequency. However (and this may be true of any filter), the rise/decay time can be dominated by the filter bandwidth instead of the source timing. Narrow filters ring. They are slow to start-up and slow to fade-out.

FFT is another route, but does not directly give octave-like bands, it is a linear process, more detail at 9 kHz than at 38 Hz.

Click to expand...

I'm pretty sure there are transformations with warped frequency scale although I've never gotten around studying them further. IIRC mp3 uses such.

A few hundred dollar for a decent spectrum analyser is surely optimistic; check what a new Agilent costs...

[Edit: Of course a full-featured analyser is probably not what you need. It is in any case helpfull if you could define what you really need:

* is real-time operation necessary
* do you need log or lin frequency resolution
* what frequency resolution do you really need (as high as possible is not a valid answer--I'm used to do FFTs with >100'000 bins)
* what amplitude accuracy do you need]

Samuel

The usual technique is cascaded down-sampling. It reduces the number of calculations. In layman's terms, if you want to have 2Hz accuracy at 20Hz (which is really the minimum acceptable resolution for 1/6 octave) with a single resolution analysis, you need a 16384-point FFT. If you down-sample in cascade, you can do 9 iterative downsamplings, each one processed by a 256 point FFT, giving enough RELATIVE resolution of ca. 1%. The MPU load goes from 16384 to 2560. Downsampling uses some CPU load but overall there is a significant benefit.
To specifically answer bodega' original post, I believe, as PRR,  a hardware solution would be huge and costly.
If I were in your shoes, I would start with an existing plugin such as this one
http://www.bluecataudio.com/Products/Category_Analysis/
and try to interface it with a specific display.
Maybe something like that:
http://www.superscreen.com/led/en/rental.php?navid=3
???

Why not use some old PC or minilaptop, put into a box and wire the screen to frontpanel? Use Linux as platform for software reliability and configurability and some suitable plugin or software analyzer. Put a trackball or similar to frontpanel for userinterface.

With a bit of luck it might last for 5-10 years and won't cost more than few hundred + case and you can use it for other metering purposes too.

If 1/24th octave would be enough, why not use Smaart? 

Link bellow has tons of info on measurements, FFT, etc..

http://www.eaw.com/products/software/resource.html

> cascaded down-sampling.
> load goes from 16384 to 2560.

Thanks. I suspected there should be something like that, but never dug into it.

> I'm pretty sure there are transformations with warped frequency scale although I've never gotten around studying them further.

If you have not gotten into them, it will be "too much" for most of us.

> define what you really need

Yes, that would be the first step. And perhaps a very loooong step to anybody who has not studied and built small spectrum analyzers already. The number of decisions can be quite overwhelming! I recall agonizing over Q decisions, scan-rates, etc. Apparently I was not the only one doing this at the time. When I had a "might work" plan, the next month Popular Electronics published a plan which echoed mine in all but a few details. (My power scheme had different aims, I used different nominal resistors because I had those values...) I built mine, I had bad LEDs (they did not matrix well), I debugged, it worked good until my assistant put the batteries in backward. So that was anOTHER lesson: defend against careless users. And build to re-build: I could have replaced all 80 LEDs, but I had made such a botch with the initial weaving and then the debugging, that I just lost heart.

Voxengo SPAN is an excellent audio analyzer, and it's free (Windows VST Plug-in).

http://www.voxengo.com/product/SPAN/
If you need a hardware analyzer, you could look up some simple VST Host and run it on a small netbook.

It's not too troubling to model a 64-band bandpass circuit in DSP, without having to get involved with the maths for FFT.  That DSP is simple enough that even I can manage it (I've done 31 band with 36dB slope, it runs reasonably as a VST on my c. 2002 PC); program it on a reasonable DSP chip coupled with a small display and the cost could be maybe in the $40 range.