Hi,

I wrote scopes for pure data via py/pyext and python, which draw waveforms and spectograms live with low cpu usage.

Drawing waveforms works well, but I'm not not shure, if the spectogram is done correctly. Have a look at the screenshots:

The default view looks ok to me, but the zoomed view shows repeating patterns.
As I'm new to analyzation with fft, I'm not shure, whether:

• it's a usual behaviour
• it's a mistake or matter of the windows function
• it's a mistake in calculating the ffts

Can someone with more experience than me help me out?

Thanks for any comments in advance,
Xaver

P.S.: The code can be found here: https://github.com/XRoemer/Pure-Data-Realtime-Scopes/blob/master/pd_scopes.py
fft processing is done in lines 609 - 620

PPS: The window I used is a hanning window (numpy) Line 880

Sorry I don't have enough experience to say what's going on in regards to texture of the sonogram, but I would think there are techniques to get much higher resolution when zooming in.

Curious, why Qt over tk?

"there are techniques to get much higher resolution when zooming in"
As I said, I'm new to fft. As far as I understand it, the resolution of the fft is half of it's window size.
So analyzing 2048 Samples result in 1024 fft values. That's what can be seen in y direction in my spectogram.

In x direction I do one fft analysis every 64 Samples. That's a little bit slower than what can be done in the spectrum example of pd.
(Every [metro 1] = 44.1 Samples)

It's more vispy over tk, as vispy is developed for scientific visualisation in real time. http://vispy.org/
Qt is only used for the window. I don't think tk would be able to do anything similar live, whithout blocking the cpu.

After two more weeks on research I found following answers:

• The strong aliases were produced by a fault in the calculation of the fft. My newer version produces less aliases.
  Screenshot
• Aliases will disappear when the fft is calculated for every sample. For a live situation that's too demanding for the cpu - at least via python an numpy. I calculate the fft every 64 samples, which is still quite demanding. But it works, when not too much other calculation has to be done.
• If I understand it right: pd calculates the fft once per window size one uses in the subwindow - so that's even much slower (Usually every 1024 or 2048 samples). I couldn't write the results of the [rfft~] faster to an array than that what a [bang~] gives me. Maybe pd calculates faster but messages are only send once per block size.

Please correct me if I'm wrong.

Edit: pd seemes to be faster when using overlapping windows. But I didn't figure out yet, what exactly is happening.