judyfong/g2p-lstm

LSTM encoder-decoder sequence-to-sequence models for Icelandic

LSTM encoder-decoder sequence-to-sequence models for Icelandic

This directory contains an LSTM encoder-decoder sequence-to-sequence models, trained for Icelandic g2p. The models were trained using the baseline for the Sigmorphon 2020 Shared task in multilingual g2p, with manually transcribed training data of ~5,800 words per pronunciation variant.

See code for training and evaluation: https://github.com/sigmorphon/2020/tree/master/task1

Reference paper: Gorman, Kyle et al. (2020): The SIGMORPHON 2020 Shared Task on Multilingual Grapheme-to-Phoneme Conversion. In: Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology (https://www.aclweb.org/anthology/2020.sigmorphon-1.2/)

Fairseq setup

1. With Conda:

conda is recommended for a reproducible environment. Once you have conda installed, create a new conda environment by running this:

conda env create -f environment.yml

The new environment is called "fairseq-lstm". Activate it by running this:

conda activate fairseq-lstm

2. Clone Fairseq and install, see: https://github.com/pytorch/fairseq