Overlap FFTs for increased time resolution?

Chet Singer
Chet Singer Member Posts: 56 Advisor
edited March 2022 in Building With Reaktor

I've been building ensembles that use additive synthesis or sample reconstruction based on FFT results. The FFT package I'm using is an old one called SNDAN and I've been analyzing wind instruments with it. There's an ensemble in the UL now that was produced using this:

https://www.native-instruments.com/en/reaktor-community/reaktor-user-library/entry/show/14437/

It works: SNDAN produces text tables of harmonic amplitudes and frequencies which I can use to drive Reaktor's Sine Bank, or to reconstruct new samples using a program I wrote.

But I'm dissatisfied with how low-pitched notes turn out, particularly during the attacks. They're lacking in detail. I think it's because the windows are wide with respect to what's really going on.

Can I use overlapping windows to reclaim these details? And how? SNDAN itself doesn't appear to overlap. Or perhaps there's another FFT tool that will do it for me.

Comments

  • ANDREW221231
    ANDREW221231 Member Posts: 295 Advisor

    increased overlaps does give something like increased time resolution for FFT, in my experience


    with respect to resynthesis with the sine bank, this means updating the parameters 2x, or 4x as often. i think you would be unlikely to regain any time resolution if the analysis program you are using does not take advantage of increased overlap


    **also, i believe from what i've read, increased overlap doesn't really help with time/frequency uncertainty tradeoff, as its sort of like a filter smearing events across time. it does help with things like latency, as in with increased overlap you will start to get back usable signal faster if using FFT in realtime

  • Chet Singer
    Chet Singer Member Posts: 56 Advisor

    Thanks. I found a paper from Tektronix that describes overlapping. I may try it.

    If I overlap 16x my results will contain 16 times the amoung of numbers. I suppose I should use all of them when regenerating the signal, so that my output windows will be 1/16 the original size.

  • ANDREW221231
    ANDREW221231 Member Posts: 295 Advisor

    i recommend it. in my experience, an overlap of around 4 is where things start to get good enough. you may want to mess around with window width as well, to taste

  • Chet Singer
    Chet Singer Member Posts: 56 Advisor

    Thanks. I will. I think window width is fixed automatically by SNDAN, though. It's optimized for tones with harmonics that are integer multiples and it asks for a root frequency, not a window width.

  • Studiowaves
    Studiowaves Member Posts: 451 Advisor
    edited April 2022

    The demo is real nice Chet. Sounds like a sampled sax. What wind controller do you use? It really seems to follow your articulations well. Um, Analog device might have a decent FFT, one of the engineers there is an old friend of mine, let me know if you can't find anything and I'll ask him if he's got one. Here is one example of a spectrum analyzer, one of many but has good explanations on how things are done. https://wiki.analog.com/university/tools/m1k/alice/sa-users-guide?rev=1611784869

  • Chet Singer
    Chet Singer Member Posts: 56 Advisor

    I used an Akai EWI Solo. It has an on-board synthesizer. But you can also use it as a controller by plugging it into a computer's USB port, which is what I did.

    Thanks for the FFT-related link.

Back To Top