Typically tutorials and articles explaining OLA (like the one on the website https://www.dsprelated.com/freebooks/sasp/Overlap_...)start with the simple use case filtering a long sequence where we choose NFFT >= Nx+Nh-1 to avoid aliasing.
This linked article next explains OLA with spectral modifications (say spectral subtraction speech enhancement) and how it can be interpreted as having a variable filter on every frame. What I couldn't understand is how do we come up with NFFT in this case? We don't know how far these spectral modifications could spread in time. How do we make sure aliasing is avoided.

I also looked at the paper referenced in the article: https://jontalle.web.engr.illinois.edu/Public/Alle...

From section V.B:
" If a modification P(eJWk) has a time response which is effectively N0 points long, the analysis length L must be at least N + N0 - 1 where the window length is N."
but then how do find out what N0 (and N + N0 - 1=L = NFFT) is?

Thanks,

Milad

 

NFFT=2*Window length

• Hi Mil1o, for spectral subtraction WOLA (weighted overlap add) is a better option because spectral subtraction is a non linear operation. In this approach you don't need to add anything to your window length. Hope this helps.