Skip to content

Latest commit

 

History

History
61 lines (55 loc) · 1.77 KB

README.md

File metadata and controls

61 lines (55 loc) · 1.77 KB
Raw

RNN-Transducer

This is a PyTorch implementation of Sequence Transduction with Recurrent Neural Networks speech recognition paper

RNN-T

Image Source: Here

@article{DBLP:journals/corr/abs-1211-3711,
  author    = {Alex Graves},
  title     = {Sequence Transduction with Recurrent Neural Networks},
  journal   = {CoRR},
  volume    = {abs/1211.3711},
  year      = {2012},
  url       = {http://arxiv.org/abs/1211.3711},
  eprinttype = {arXiv},
  eprint    = {1211.3711},
  timestamp = {Mon, 13 Aug 2018 16:48:55 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1211-3711.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Train on your data

In order to train the model on your data follow the steps below

1. data preprocessing

  • prepare your data and make sure the data is formatted in an CSV format as below
audio_path,text,duration
file/to/file.wav,the text in that file,3.2
  • make sure the audios are MONO if not make the proper conversion to meet this condition

2. Setup development environment

  • create environment
python -m venv env
  • activate the environment
source env/bin/activate
  • install the required dependencies
pip install -r requirements.txt

3. Training

  • update the config file if needed
  • train the model
    • from scratch
    python train.py
    • from checkpoint
    python train.py checkpoint=path/to/checkpoint tokenizer.tokenizer_file=path/to/tokenizer.json

TODO

  • adding the inference module
  • Adding Demo