Excellent, thanks. One reason I'm wondering is that many definitions of window functions in DSP texts and online define Hann as something like 0.5*(1 - cos((2*pi*n)/(N - 1))), which is the symmetric version, 0 at both ends. With overlap-add it puts a repeated value at both peak and trough, but still with the FFT periodicity.

On Sat, Jan 9, 2016 at 7:02 AM, katja <katjavetter@gmail.com> wrote:

Hi Matt,

What I remember from FFT experiments is, a 'periodic' Hann window
(asymmetric) gives best results for overlap-add. You can think of the
window as two functions: the DC component preserves the original
signal, while the cosine component produces sum- and difference
frequencies which cancel each other in subsequent overlaps after IFFT.
Long story and illustrations here:

http://www.katjaas.nl/FFTwindow/FFTwindow&filtering.html

Katja

On Sat, Jan 9, 2016 at 6:39 AM, Matt Barber <brbrofsvl@gmail.com> wrote:
> Hi list, quick DSP question.
>
> In I02.Hann.window.pd there is a Hann window computed by running an [osc~]
> for one cycle over 512 samples. Doing it this way is simple but it means
> that it doesn't taper all the way to 0 on the right (the next value of the
> oscillator after writing the table is 0). I don't know why I hadn't thought
> of this before, but is this asymmetry mathematically optimal for windowing
> of a single frame? What about in an overlap-add context? I'm trying to
> understand the logic, so I'm not too worried about whether it makes a
> practical difference. A period of 511 samples would be symmetric and
> 0-valued at both ends, but there would be no "middle" value at unity.
>
> What's the best way to think of this?
>
> Matt
>

> _______________________________________________
> Pd-list@lists.iem.at mailing list
> UNSUBSCRIBE and account-management ->
> http://lists.puredata.info/listinfo/pd-list
>