🚀 go-pugleaf

Hello....
I'm applying various bilinear transforms/distributions (such as
Wigner, Choi-Williams, etc) to speech and am pondering some of the 
rarely mentioned practical details.  

Specifically, bilinear transforms take a window of time domain data,
multiply the data points' amplitudes together (with various combinations
and scalings of the amplitudes of the data points depending on the 
particular bilinear kernel used), and then typically take the FFT
of the newly constructed data to provide the Time-Freq distribution.

How should the original time domain signal points be scaled and offset
wrt 0.0 amplitude such that the bilinear tranform's initial multiplication 
of data points in the time domain has the optimum effect in extracting 
useful spectral information in the final T-F output distribution?  

1) Intuitively and empirically, setting the offset so that all time 
   domain data is positive before doing the multiply seems to badly 
   distort the output spectrum.
2) Setting the offset to remove the root of the DC over the window does 
   not work as well as 3 below.
3) Setting the offset to remove DC over the window seems to work fairly 
   well..... but is there something better?

Ideas? Thanks.
Lauren Hodgson
Dept of Systems Design Engineering
University of Waterloo