From a machine learning perspective this doesn't seem like such a daunting task. We have pretty good speech recognition for human language (plus acceptable speech synthesis). Assuming dolphins have a smaller vocabulary than humans, speech recognition of dolphins should strictly speaking be a simpler problem (although it would be significantly harder to record the training data).
Oh, I'm sure there would be a few unfortunate "drop your panties, Sir William; I cannot wait until lunchtime" incidents, but with an automated Hungarian Phrasebook[1], we'll eventually learn to get by.
More to the point, though, we have no reason to assume that dolphins, any more than, say, apes, have the ability to "chat" -- to communicate about things displaced from the here and now. It's worth a go, of course, but our expectations shouldn't be too high going in.
I think part of the problem may be context. Without swimming around with them, and having similar senses (something that these days some wearable sensors might mitigate), it may be difficult to grasp what they are talking about.
