I still have a lot of open questions (see message board), but I rethought the approach to what you need, and it should be possible, possibly within my thesis (which can be kept secret for a period of time where I live) and most certainly in collaboration with able colleagues of mine.
It would be a quite cpu intense neural networking application, cluster based by design.
Also, there would have to be paid milestones, e.g. for a certain recognition rate for certain words.
It is a great challenge to push the limits of whats possible today, but a fun one at that :)