I'm using RDRPosTagger[1], though I've optimized the code a bit so that it's not just algorithmically efficient, but to use the language in a way that is fast. It isn't perfect, but it's good enough to be useful.
Language detection and sentence splitting are the other two slow bits of processing.
