1
$\begingroup$

I have two filters which must apply (in series) to same input data $x[n]$ e.g of length 256 (real time asio input buffers). One filter is e.g. an impulse response $h_1[n]$ of a speaker and other one an IR $h_2[n]$ of a lowpass (created by sinc, blackman). For simplicity lets say both are shorter than 256.

So I zero pad the input and both IRs to 512, FFT and then do: $X[k]\cdot H_1[k]\cdot H_2[k]$, after this I do iFFT and use overlap-add to deal with the output vector of length 512. I do use a wrapped version of FFTW for performance reasons. FFTW tutorial says that doing FFT and afterwards iFFT of a vector of length $N$ produces an output vector which must be scaled by $\frac{1}{N}$.

My understanding problem here is, how do I have to scale? In the upper example I don't just do FFT$\to$iFFT, but I am doing FEW forward FFT of different filters, multiply them and do only one backward iFFT.

Improve this question
$\endgroup$
4
  • $\begingroup$ the iFFT in FFTW does not do the conventional scaling of $\frac1N$ that is in the most common convention of the definition of the inverse (or "backward") Discrete Fourier Transform. I try, in this answer, to be explicit about the consequences (regarding convolution) of convention of where the $\frac1N$ scaling factor goes. $\endgroup$
    – robert bristow-johnson
    Commented Aug 6, 2019 at 7:49
  • $\begingroup$ I am not sure if I get it right. If I use FFTW for e.g. 1 input x[n] and three filters h1[n], h2[n] and h3[n] all zero padded to N, then do FFT on each and multiply those three Y[k]=X[k]*H1[k]*H2[k]*H3[k]. After this I transform back using y[n]=iFFT[(Z[k]). How should i skale y[n]? $\endgroup$
    – f.f
    Commented Aug 6, 2019 at 8:16
  • $\begingroup$ by $\frac1N$ where $N$ is the length of the FFT and iFFT in FFTW. also, keep in mind that the linear convolution of a sequence having non-zero length $M$ with another having non-zero length of $L$ results in a sequence of length $M+L-1$. so when convolving four sequences together, add the non-zero lengths of them all. the FFT length $N$ must be as large as that sum of lengths. zero pad to that length $N$. $\endgroup$
    – robert bristow-johnson
    Commented Aug 6, 2019 at 9:04
  • $\begingroup$ so before I added the 2nd filter I had: signal x(N1=256), filter h1(N2=256) and zero padded each to N=512 before fft. all worked fine, Now I have an additional filter (which I multiply in freq domain) of h2[L] with L (0<L<256). I zero padded this one also to 512 and went on. This was wrong right? since I have N1+N2+L which is >512 I need to zero pad now all inputs x,h1 and h2 minimum to length N1+N2+L right?? Could this explain my strange sizzle after adding the filter h2. Aliasing? $\endgroup$
    – f.f
    Commented Aug 6, 2019 at 9:51

0

Reset to default

Your Answer

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

Browse other questions tagged or ask your own question.