Dictionaries v. LTS models

[Sandy F]

As there are already some widely-used phonetic dictionaries in existence, why wouldn’t it make sense to add more proper names and unusual words to these dictionaries rather than use LTS models? I understand that there will always be words not included in any dictionary (new creations, really uncommon words, etc.), but are LTS models fundamentally more efficient than storing more words in a dictionary?

[Simon]

Adding more proper names would be a good idea, because those are generally hard to get right with LTS. But we could never cover every possible proper name. Likewise, adding “unusual” (i.e., low frequency) words cannot in itself solve the problem.

One simple reason is that new words are invented all the time, and no dictionary can include every possible word we might encounter.

Your comment about the efficiency of LTS is spot-on though, in terms of storage space. After creating the LTS model, we could remove all words from the dictionary that this LTS model correctly predicts. That would reduce the size of the dictionary. Festival does not do this, but commercial systems may.

[Author]

Posts