I understand that DFT convolution of a signal with a filter kernel results in circular convolution. Techniques such as Overlap and Add and Overlap and Save take care of the corrupt samples and produce a clean signal.
Now, what if you DON'T convolve the signal with anything and simply manipulate the bins directly? For example, in a "Spline EQ". How are circular convolution artifacts eliminated at the frame boundaries when doing direct bin manipulation like this?
FFT convolution vs. direct bin manipulation
-
- KVRian
- Topic Starter
- 625 posts since 30 Aug, 2012
Hmmm - no one knows - or do you not want to tell us?
There are lots of DFT-based processors around that are not simply convolution filters - they perform advanced processing and manipulate the bin data directly. How are iDFT circular artifacts avoided in this case? Standard 50% OLA and OLS methods don't work (I've tried them).
There are lots of DFT-based processors around that are not simply convolution filters - they perform advanced processing and manipulate the bin data directly. How are iDFT circular artifacts avoided in this case? Standard 50% OLA and OLS methods don't work (I've tried them).
-
- KVRian
- 1379 posts since 26 Apr, 2004 from UK
As soon as you are manipulating the spectrum, you are doing some kind of a convolution, so the same rules applies. The issue is that some manipulation can have an impulse length bigger than your original signal.