The slide below, borrowed from CMU, shows the typical interpretation of the "filterbank interpretation" of signal reconstruction from a Short Time Fourier Transform. As far as I can tell, the hop-size here is 1, i.e., there is no frame decimation to represent a realistic frame hop.

In that context, it makes sense: your inputs for any given sample $n$ are the outputs of lowpass-and-downconvert operations, so to reconstruct, you just upconvert-and-add them together. Do this at each timestep.

But usually, we have a substantial hop, which means that the $X[n,k]$ elements only exist at hopped values of $n$. I.e., if our hop is 1, we have $X[0,k], X[16,k], \ldots$

Consequently, we are only calculating $y[0], y[16], \ldots$

How exactly do we get the missing values? Interpolate? If interpolation, where does that happen? In the branches after the upconversion? At the $y[n]$ output?

I have scoured multiple articles, slidesets, and books, and none of them seem to explain this.

Or am I mis-interpreting what's happening here? (If it matters, I'm interested in audio, where window sizes would be a few hundred samples, and the hops would be 25% of that.)

enter image description here


1 Answer 1


Here is (hopefully) what you're looking for: https://ccrma.stanford.edu/~jos/sasp/Downsampled_STFT_Filter_Banks.html.

Specifically, this section: https://ccrma.stanford.edu/~jos/sasp/Filter_Bank_Reconstruction.html

  • $\begingroup$ It definitely points me in the right direction and partly confirms my intuition. I need to ponder that. $\endgroup$
    – Novak
    Aug 11, 2022 at 6:48
  • $\begingroup$ Glad I could help. Let me know if anything is unclear ;) $\endgroup$
    – Jdip
    Aug 11, 2022 at 7:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.