There's good voice-recognition software that supposedly needs about 15MIPS (ViaVoice) to run in real-time, and you may not even need real-time. I don't know how much bandwidth to how much memory speech-recognition needs, though.
Looks like a fair amount: ViaVoice not only specs system memory (64-96M) but also 256K of L2 cache. There are some people nearby working on automatic segmentation/speaker recognition/tagging and search, etc., and they use serious crunch for their applications, but maybe they'd have some ideas for what sorts of things could be done cheaply.
-Dave
