There is a type of natural selection pressure on language, mentioned above, where language items that are understandable and replicable and transmittable and memorable are passed on while others are not - I think there's a weak selective pressure for this (language items that are already in use have already gone through this filter - it would probably apply more to neologisms)
Much of the contemporary research in phonological theory is heading in these directions. The most promising new work right now is probably Blevins's book
Evolutionary Phonology (2008). The argument there takes evolution as a lot more than a metaphor, though for her it's phonological systems rather than cognitive faculties that are evolving.
The idea is basically that articulation and perception are inherently tied to noisy material channels, and thus the sounds encountered by a speaker of a language intrinsically include a broad scope of quasi-random variation. This scope of variation converges over time on relatively more stable, more contrastively adaptive type-categories in highly patterned (though completely atelic) ways, and tracking these pattens allows us to explain both the synchronic system and the diachronic change of languages simultaneously.