I see a potential problem with any phase based deterministic wallet. It reduces the namespace of an attacker trying to force an address collision by searching for English phrases in the same way that a dictionary attack works against common passwords. It's more than conceivable to have an accidental collision as well, if two fans of classical lit both choose "Call me Ishmael" or more than one Tolkien fan chooses the same quote from TLOTR. Hell, an attacker who was just using the King James version of the Bible would get quite a few hits from Christians using their favorite verses. It would be better to do it in reverse, by having the client generate a random number sequence and translating that into a set of English words that can be printed, saved as an encrypted file to be stored elsewhere, or memorized.
Electrum doesn't work that way it uses a predefined list of words based on a key but still even if the passphrase is freeform there are ways to overcome that.
The first is used of salt. Including a non-secure semi-unique value in the key generation process like user's email address. This doesn't need to be secure but it should be semi-unique. This prevents using a pre-computation attack as each user's hash is unique even w/ same passphrase.
I was thinking more along these lines, and wondering if a passphrase plus a salt created by a standardized questionaire of usually secret personal info could be used, of the kind of questions that don't change. For example, one such question could be "How old were you when you lost your virginity?" with multiple choice answers including each age from 12 to 24, and an option like "does not apply/refuse to answer" so that the multiple choice questionaire could take all such answers, as well as the numbers of the answers that users refuse to answer, and create a salt that could produce a unique. The questions would have to be high in number, and of a standardized order so that a user could concievablely reproduce the wallet.dat while be unique enough that it won't produce wallets that could collide. It would have a bias, as all such questionaires do, but it should a long enough of a questionaire that such a bias isn't predictable and of such personal info that users aren't going to answer such a questionaire outside of the context of the client.
