On Wikipedia there is quite a list of window functions. However, when you are doing a lot of STFTs, you often have to have some overlap to get better results.
Generally an overlap factor of 2 (i.e., 50% overlap) works fine (I'm doing analysis & resynthesis), but for some windows it isn't good (I currently have this experience with the Blackman window).
I tried my best but I couldn't find a resource that would list the "good" overlap factors for common and less common windows.
Please refer to this paper on the optimum overlap percentage for the Blackman-Harris window, which is derived to be 66.1%. It has a lot of useful information on spectral analysis and windows.
I tried my best but I couldn't find a resource that would list the "good" overlap factors for common and less common windows.
Here's a list of window functions and overlap factors that have constant overlap-add (COLA). (Code here)
Heinzel - Spectrum and spectral density estimation... shows several windows and lists their "optimum overlap" and their "amplitude flatness" (AF graph) for different amounts of overlap. The overlap-add is flat when the AF graph is 1.0, so for Hann, for instance, it's COLA at 50%, 66.66%, 75%, etc.:
Their "optimum" overlap is the best trade-off between amplitude flatness and wasted computation time (not necessarily COLA). (The Hamming window's AF curve looks like it doesn't reach 1.0, but it actually does at tiny points if you zoom in.)
Note that I've listed these as fractions of overlap, not percent, since that's more natural. The COLA overlap that has the best trade-off is shown in bold.
boxcar()rectwin()barthann()barthannwin() bartlett()triang
triang()blackman()blackmanharris()flattop()flattopwin()hamming()hann() nuttall()nuttallwin() parzen()parzenwin()tukey(alpha=0.5)tukeywin(r=0.5) Also, Borβ-Martin points out that you can generate a COLA window for any arbitrary overlap by convolving an overlap-length window of area 1 with a hop-length rectangular window. They provide a particular family of windows that let you trade off main lobe width and sidelobe fall-off rate.
The area under the curve of a Blackman window is less than 50% that of a rectangular window (for a von Hann window the percentage is exactly 50%). So more information will be lost (to quantization or other noise) by the windowing process using a Blackman window than using a von Hann window (or using a rectangular window for that matter). Using overlapping windows provides some information redundancy (the same samples will be processed multiple times into the results), so more overlap can make up for some of this loss of information.
Using more overlap provides more time locality resolution as well as more informational redundancy. However, since time resolution is also affected (blurred) by the effective window width, there will diminishing returns in using too much overlap (e.g. stop increasing overlap when your analysis process no longer sees "useful" differences between offset frames).
50% overlapped von Hann windows also "add up" to a flat top window, which is useful for resynthesis without amplitude modulation distortion, no matter if a different window is used for analysis.
Maybe look for the keywords ,,shift overlap add property'' or ,,constant overlap add property'' - Some window functions have the advantage to sum up to a constant function if taken in the proper distance, as was remarked here.
