Wav2Vec2 ASR

This repository contains a implementation of Wav2Vec2 and also provide codes for fine-tuning on Wav2Vec2-base of Vietnamese.

Demo is available at Hugging Face

Prepare Dataset

The code is fine-tunned on VLSP2020 Dataset
Since I mainly train on Google Colab, so I need to convert this dataset to Web Dataset for faster loading from Google Drive

cd finetuning
python preprocess.py --data_dir [DATA_DIR] --dest_dir [DEST_DIR]

where:

• DATA_DIR: Path of the VLSP2020 data
• DEST_DIR: The directory where data is extracted in

Fine-tuning

You can view my fine-tuning in this notebook and WandB

cd finetuning
python train.py \
    --batch_size 2 \
    --num_workers 2 \
    --classifier_lr 1e-4 \
    --wav2vec2_lr 1e-5 \
    --max_epochs 3 \
    --accelerator gpu
    --grad_clip 1.0
    --data_dir [DATA_DIR]
    --ckpt_dir [CKPT_DIR]

where:

• DATA_DIR: The directory that contains extracted data
• CKPT_DIR: The directory that checkpoint would be saved

Inference

Please view this notebook