Later in the program, Holland says, PAL will be able to "automatically watch a conversation between two people and, using natural-language processing, figure out what are the tasks they agreed upon."
PAL's role here is not clear. The 'easier' task would be to monitor the body language of the two conversers and, by lining up a list of tasks with the observation of their head movements, correctly predict which points in the conversation were the ones where someone performed an "agreement" gesture.
The much, much more difficult task would be to actually read lips. There are only certain properties of phonemes you can deduce from how the lips and jaw move; many, many other features of speech are lost. Only when you supply the machine with a limited set of words in a limited topic domain do you get good performance; otherwise, you're grasping at straws. And then taking out most of the speech signal? Please.
But no, DARPA is cool and will save all the translators in Iraq (by 2009, well before the war ends.) PR and vaporware win the day!