This is a set of high-level utilities for building data sets for multilingual g2p systems. It also includes support for training and preprocessing those models using OpenNMT, and for computing g2p error metrics on the test set. I used it to create the models for my paper for the Workshop on Building Linguistically Generalizable NLP Systems at EMNLP 2017 (paper preprint here).

The basic usage is like this:

python mg2p.py spanish-model -preprocess 

Where the first argument is the name of the model you want to create. The flags -preprocess, -train, -test, or any combination of these three may follow, depending on how much of the process you want to do in a single command. Other optional arguments can specify which languages and scripts to include in the training/validation/test data and what features to append to the source and target.

Pronunciation data for the models is available here.

NOTE: Due to (over)zealous adoption of language identification feature embeddings in the months since I wrote the paper, mg2p.py does not actually currently support the language identification token approach described in the paper. This will be rectified soon.