ASR project barebones

Report

About all details about this homework can be found in wandb report.

This is a repository containing a convenient pipeline for training automatic speech recognition models.

Advantages of this repo:

possibility of changing experimental configuration by only tuning one json file
good and clean code structure (see hw_asr folder with all elements of pipeline)
prepared scripts for training and evaluation of models
prepared downloadable checkpoint with high perfomance on librispeech test sets

To set up the environment for this repository run the following command in your terminal (with your virtual environment activated):

pip install -r ./requirements.txt

Make sure all tests work without errors

python -m unittest discover hw_asr/tests

To download my best checkpoint run the following:

python default_test_model/download_best_ckpt.py

if you are interested how I got this checkpoint, you can read about that in wandb report.

You can evaluate model using test.py script. Here is an example of command to run my best checkpoint with default test config:

python test.py \
  -c default_test_model/config.json \
  -r default_test_model/checkpoint.pth \
  -t test_data \
  -o test_result.json

Use train.py for training. Example of command to launch training from scratch:

python train.py -c hw_asr/configs/config_librispeech.json

To fine-tune your checkpoint you can use option -r to pass path to the checkpoint file:

python train.py \
  -c hw_asr/configs/config_librispeech.json \
  -r saved/models/<exp name>/<run name>/checkpoint.pth

This repository is based on a asr_project_template repository.