I prepare a text for the TTS Dataset to be spoken. Just woundering, should I dismiss all abbreviations? What if I leave some of them, let’s say, “USSR”, will it learn to pronounce exact that word (well, in the lower case) letter by letter? Or it most likely gonna cause training and inference difficulties?
Thank you for your suggestions!
If it’s an english dataset, if you leave abbreviations as is, it’s unlikely the tts is going to infer it right( because of how proportionately low those occurrences are compared to normal words ).
You could definitely train the tts, and abbreviations would probably be brushed over as noise.
Please correct me if I am wrong.
If you use phonemizer and if the rule is defined for that abbreviation in it then it should be fine. Otherwise, it’ll do its best to approximate. So check how phonemizer behaves for your abbreviations.
If you use characters, it will probably just spell the given abbre.
I’ve already just checked, phonemizer has no ability (as long as I didn’t modify it) to generate other than general word transcription for the abbriviations. Seems like with non-tuned phonemizer its better to remove all sentences with abbriviations, which I gonna do now.
Thank you everyone for the answers!