How to create weighted FST for Hebrew -> IPA

[thewh1teagle]

Hey!
I’m looking into using rustfst for a Hebrew grapheme-to-phoneme task. The input is Hebrew text with diacritics and punctuation, and the output should be the IPA phoneme sequence. Non-Hebrew characters (like punctuation or English words) should just pass through unchanged.

For training I have a dictionary with Hebrew words mapped to their IPA. What’s the best way to create a weighted FST for this? Should I try to generate it directly in Rust from the dictionary, or train it with an external tool like Phonetisaurus and then load the .fst into Rust?

[Garvys]

Hello @thewh1teagle !
You can directly create an FST with rustfst by using a VectorFst.
Then you need to add states with the add_state() method, add transition with the add_tr method, set a state a final with the set_start() method and set a state as final with the set_final() method.
Hope this helps !

[thewh1teagle]

@Garvys
Thanks! I didn't actually understand how to create weighted FST so I just went ahead and created FST in Python.
I'm wondering if I can just load the compiled model and use it with Rust easily, the input is just word (like from the tests) and output is phonemes

• https://github.com/thewh1teagle/heb-fst