README.md

Capitalisation And PuncTuatION (CAPTION)

PT2020 Transcription project.

In this repository, we explore different strategies for automatic transcription enrichment for ASR data which includes tasks such as automatic capitalization (truecasing) and punctuation recovery.

Download IWSLT corpus

Publications:

• Multilingual Simultaneous Sentence End and Punctuation Prediction
• Towards better subtitles: A multilingual approach for punctuation restoration of speech transcripts
• Automatic truecasing of video subtitles using BERT: a multilingual adaptable approach

Sentence end and punctuation prediction shared task

To replicate our winning submission to SEPP 2021 please go to the shared-task branch.

Model architecture:

Available Encoders:

• BERT
• RoBERTa
• XLM-RoBERTa

Requirements:

This project uses Python >3.6

Create a virtual env with (outside the project folder):

virtualenv -p python3.6 caption-env

Activate venv:

source caption-env/bin/activate

Finally, run:

python setup.py install

If you wish to make changes into the code run:

pip install -r requirements.txt
pip install -e .

Getting Started:

Train:

python caption train -f {your_config_file}.yaml

Testing:

python caption test \
    --checkpoint=some/path/to/your/checkpoint.ckpt \
    --test_csv=path/to/your/testset.csv

Tensorboard:

Launch tensorboard with:

tensorboard --logdir="experiments/lightning_logs/"

If you are running experiments in a remote server you can forward your localhost to the server localhost..

How to run the tests:

In order to run the toolkit tests you must run the following command:

cd tests
python -m unittest

Code Style:

To make sure all the code follows the same style we use Black.