I have trained a FastPitch model with a custom dictionary that has entries of the following format:
grapheme phoneme
Feedback ˈfiːtbɛk
italienische ˌiˑtalˈi̯eːnɪʃə
The trained NeMo model picks up the entries from the dictionary properly but once deployed to Riva this is not the case anymore. I figured out that this happens because Riva expects the dictionary entries to be all uppercased. Hence a workaround is to manually uppercase all left hand side entries of the dictionary file and pass the updated dictionary via --phone_dictionary_file during riva-build.
I have trained a FastPitch model with a custom dictionary that has entries of the following format:
The trained NeMo model picks up the entries from the dictionary properly but once deployed to Riva this is not the case anymore. I figured out that this happens because Riva expects the dictionary entries to be all uppercased. Hence a workaround is to manually uppercase all left hand side entries of the dictionary file and pass the updated dictionary via
--phone_dictionary_fileduringriva-build.