Any inference tricks to preserve abbreviations (e.g. S.O.S.)

[schlumpf007]

I am using the model to create mock conversations.

Is there a way to preserve abbreviations so they are spoken letter by letter?
E.g. when trying "SOS is a serious signal" -> it will not say "es oh es ..." but speak is like "sauce"

[SWivid]

could use S.O.S. is a serious signal

[schlumpf007]

Yes I tried, also S O S or S - O - S. Or even Es Oh Es or similar. With this particular example it gets better or worse but it's not stably reproducible with other acronyms.
Some models support custom dictionaries or phonemes. Is there any other way than finetuning? I have tried finetuning to do code mixing which works actually nice with en/german. But I would like to avoid finetuning on every acronym
Thanks for helping

[SWivid]

A. B. C. or A.B.C is stabler than others though still not very. And the prompt wav may have impact (different timbre vary a lot in read pattern), not sure if using a prompt wav that contains acronym pronunciation would help guide inference.

Some models support custom dictionaries or phonemes.

Yes, best to pretrain with mixed chars/phonemes.