phonetic_transcriber

The phonetic transcriber is a tool for converting a text to its representation in phonemes. The transcriber is based on Recurrent Neural Networks (RNNs).

Models Accuracy

Accuracy obtained from train and evaluation sets during models training.

Usage

Since we use relative paths for modules, you will need to set PYTHONPATH environment variable:

export PYTHONPATH="${PYTHONPATH}:/path/to/repository/"

The tool currently provides two features:

1. Train models: actions/train.py

   Launch the training of the phonetic transcriber using RNNs

   python actions/train.py <num_epochs>
   

   Launch the training of the phonetic transcriber using RNNs, with an attention decoder

   python actions/train.py <num_epochs> --att
   

2. Evaluate models: actions/evaluation.py

   Launch the evaluation of the phonetic transcriber with RNN models

   python actions/evaluate.py <encoder_model_path> <decoder_model_path>
   

   Launch the evaluation of the phonetic transcriber with RNN models, with an attention decoder

   python actions/evaluate.py <encoder_model_path> <decoder_model_path> --att
   

Display available option:

python3 actions/train.py --help

usage: Train [-h] [--att] num_epochs

Training of RNNs to perform text-to-phonemes transcription.

positional arguments:
  num_epochs  Number of training iterations.

options:
  -h, --help  show this help message and exit
  --att       Whether to use an attention decoder or a plain RNN decoder.

Requirements (tested versions)

pytest 7.2.1
torch 1.13.1
numpy 1.24.2
matplotlib 3.6.3