I'm really confused by the terminology here.

A "correlation" tells how much alike two series of numbers (of the
same length) are.

You have 3 dimensions you're looking at: space (voxelnum),t, and subject. It
sounds like you want to, for each voxel, extract the time course. So, say
at voxel 423, you get a single time series for each subject. Now you want
to get a measure that says how alike all 9 (one for each subject) of your
time series are at that voxel.

Is that accurate?

If so, I don't know of a statistic off the top of my head that will do that, that is,
that will measure how alike 9 time series are.

But if I've explained your question correctly, somebody else can probably
help you.

rich