Riva expects TTS G2P dictionary to be uppercased while NeMo doesn't

[sch0ngut]

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.