README.md

ASR HW 1

Automatic Speech Recognition task using Conformer model trained on Librispeech.

Wandb Report

Installation guide

pip install -r ./requirements.txt

Download 3-gram.arpa and vocab from https://www.openslr.org/11/. To use them change kenlm_path and vocab_path in config.json. You can download model checkpoint from this link. Config for the best model can be found in hw_asr/config.json.

Training

python train.py -c CONFIG

Check hw_asr for config examples

Testing

One can choose which metrics to evaluate in config file.

python test.py -c CONFIG -r CHECKPOINT

Eventually, you will have a file with predictions, metrics are outputted to the console.