(Excuse any ignorance of py package distribution, or anything else…)
It seems a bit weird to me that the GH TTS source repo includes the models, and requires cloning to train a new model.
I’m just curious whether the models could be separated out into their own repos, with dependencies on the source, if required (or some other packaged/binary implementation… .whl?).
I think such a separation would open up possibilities for training and redistributing model packs, with far less burden on users (I personally only want to train a new model, and anticipate no need to modify source).
Perhaps an arbitrary number of models could be included in any given model repo (e.g. a default package with the original 20), and a template/default/example model repo could be created that is completely laid out to allow users to clone it, modify it as necessary to train up a new model, and submit to their own branches, with minimal source. A nice, simple model-only repo like this would be easier to follow, modify and branch, IMO.