I would always suggest to use a 'normal' function for the background/baseline,
e.g. use a cubic function for the background at the same time you are
fitting peaks with Lorentz(or Voigt) as they have long tails. If you have multiple peaks/bands
overlapping, the background might appear bigger than it actually is,
and a baseline spline might 'cut away' actual intensity..
It would be
nice to have a possibility to substract the function used for the
background after fitting. Maybe make the first function of each data set
to be the function describing the background by default. It would also
be useful to be able to just fit this background function, if using a
set of fitted functions copied from the last data set (fit won't work,
if background levels are too different).
For further reading, I can recommend Meier (2005) On art and science in curve-fitting vibrational spectra (
http://www.sciencedirect.com/science/article/pii/S0924203105000494).