• Python 2.7
• Keras 2.0.8+ with Tensorflow backend

This repository contains code, models and data for the task of English-to-Hindi transliteration

• Datasets contains the dataset for transliteration. This was obtained from the Fire 2013 track on transliterated search, more specifically from here.
  • Bollywood_dataset.txt contains the above data
  • Noisy_bollywood.txt is an artificial-noise injected dataset by random mutation, removal or addition of characters to the English words
  • Bollywood_oversampled.txt contains oversampled data, where the different English transliterations of each Hindi word are sampled equally.
• eng_hin_simple has code for training and testing
  • data.py has helper functions for loading data
  • prog.py has the training and testing code, that trains on Bollywood_dataset.txt by default, and can be modified to train on any other datasets.
  • model.png has the diagram of the model
  • The best model is saved to saved_model.hdf5
• try.py is the inferencing script, to transliterate a set of english words from a file (path to be set inside script)

• typing_noise.py is an experimental script to model noise in spellings of words, based on QWERTY keyboards' proximity of character pairs