Closed lenzo-ka closed 2 years ago
and there's no need to assume the file names are uppercase only
When sphinxtrain creates tree files with phone names on a filesystem with case-insensitive names (Windows), we have troubles.
Shouldn't that be taken care of by the client? The assumptions are platform specific. In particular, making the phone names uppercase prevents referencing the phonemes by name or the words by name. If a user wants to make a case insensitive set of names, they should simply make sure that the names don't have conflicting upper and lower case names, and then normalize the dictionary, language model, and phone set before training.
I'll put it behind an option the config
Cleaning this up and re-submitting from private branch
Uppercasing the phone names by default prevents phonesets like XSAMPA from being used, and there's no need to assume the file names are uppercase only. Also, emit a little more info while building models.