Extension of the HuggingFace Transformers for Automatic Speech Recognition

This repository is still under active development. Proper README and documentation will be added soon.

Setup

1. BUT SGE Cluster setup
   1. Clone repository and change directory to project root.
   2. Set ENVS_ROOT.

      ENVS_ROOT=/mnt/matylda5/ipoloka/envs

   3. Create conda environment

      conda create -p "${ENVS_ROOT}/huggingface_asr" python=3.10

   4. Activate conda environment

      conda activate "${ENVS_ROOT}/huggingface_asr"

   5. Install requirements

      pip install -r requirements_BUT_cluster.txt

   6. Extend PYTHONPATH with sources root

      export PYTHONPATH="${PYTHONPATH}:${PWD}/src"

   7. Run following command and test if help message was printed.

      python src/trainers/train_enc_dec_asr.py -h

Training

Recipes are provided in the recipes directory. Each recipe contains a run.sh script that can be used to train a model. The script contains all the necessary commands to train a model. The script can be run directly or it can be submitted to the SLURM cluster.