For speech recognition Pocketsphinx relies upon the existence of a function that can map from graphemes to phonemes, in other words function that can map from a written word to its phonetic spelling. Currently Pocketsphinx has a dictionary based grapheme to phoneme function. It stores all possible words and their phonetic spellings in a file, loads this file in to a map, then relies upon this map to implement the required function. This is sub-optimal in that words not in the dictionary can not be recognized. This bug's task is to replace the dictionary based grapheme to phoneme function by a WFST, a Weighted Finite State Transducer, learned from the data in the dictionary file. This WFST will then be able to produce phonetic spellings for unseen words using the rules learned from the dictionary file.

No wfst approach will be used in pocketsphinx, but a more efficient and compact way.

Aaron, this bug is invalid. Follows the right one for this feature: https://bugzilla.mozilla.org/show_bug.cgi?id=1180113