This depends on your particular needs, but I wouldn't overlook Perl's power as a glue language. For example, Matlab (or its free equivalent Octave) has a rich set of built-in linear algebra and signal-processing functions, and is the lingua franca of machine learning. I've often found it easier to talk to Octave over a pipe, via script files, and/or by reading and writing Octave-format text data than to rewrite code in PDL, or even to use Inline::Octave. I'm guessing things are similar in other domains.
in reply to recommendations on scientific computing with Perl