Fine-tuning a Wav2Vec2 model without reliance on Huggingface's Trainer.
base folder contains a list of .txt and .wav files, in which a .wav file's transcript should be contained in the .txt file with the same name.
main.sh bash script to fine-tune the model.
eval.sh bash script to evaluate the trained checkpoints on an evaluation dataset.
By default, this repository uses a Vietnamese checkpoint as in src/train_util.py. Feel free to change the original checkpoint depending on the language/model to be fine-tuned.