Consider (de)modulating stft

Most STFT implementations, including this one, do not compute STFT per standard formulation. Some use a “(de)modulated” variant, as in this paper:

This centers the windows at zero-phase of DFT bases. librosa’s implementation already has a shift (as do others), but it’s much less favorable in one regard: numeric stability. For localized component / band analysis, results are extremely sensitive to inclusion or exclusion of rows, which has adverse effects on e.g. reconstruction and synchrosqueezing. Details here.

The tweak is quite simple - consider adding as option.

[OverLordGoldDragon]

shifting only the window does not achieve the desired outcome.

Right, didn’t mean to imply otherwise.

[bmcfee]

Circling back on this; #1351 includes a fix for #1375 , in which we found an implementation error in util.frame when the input has ndim>2. The rewrite in #1351 avoids copying in all cases, and creates a read-only view of the data. This ultimately precludes anything like a “modulated” parameter to frame.

If we implement this feature, it will have to be done as a post-processing step on the stft, as I’ve advocated previously.