Separate engine from linguistic data and come up with a set of conventions for training models

[ftyers]

At the moment parts of the data are mixed with parts of the code. It would be good to have full separation of data and code and a set of conventions for formatting/organising the data such that training DeepSpeech would be as easy as "get a directory of data in the right format and execute something like: DeepSpeech.py --train ../commonvoice-fra or DeepSpeech.py --train ../commonvoice-chv --transfer-from ../commonvoice-eng .

This is not a call to immediate action, but more of a placeholder for future work.