I doubt they use bloom filters (or any other hash-based algorithm) because hashes by their design are a quick way to reference something well defined. A gesture is anything but and you need something a lot more fuzzy and flexible than hashes.

I'd think that they try to calculate the focus of each direction change and see what letter it landed on, then try a fuzzy match to words that have the smallest hamming distance between their input. After you get a good approximation of all the letters, then you can try hashing techniques.

Also, pivoting on first and last letter of each word isn't fool-proof, as the first and last letters are also not guaranteed to be hit dead-on.