SABER-labs / SABERv2

SABER - Semi-Supervised Audio Baseline for Easy Reproduction

SABER - Semi-Supervised Audio Baseline for Easy Reproduction

Easily reproducible baselines for automatic speech recognition using semi-supervised contrastive learning.

Data Preparation

• Download CommonVoice English Dataset
• Setup config.toml to use the paths where data was downloaded.
• Install requirements using pip3 install -r requirements.txt
• Prepare data using python3 -m dataset.prepare

Train

• Train using python3 -m train

Logging

• Start tensorboard using tensorboard --logdir training_artifacts/tb_logs

TODOS

• supervised training and dataset
• Check online evaluator piece from Pybolts Simclr
• Add more logs.
• streaming convnets model
• save and load projection weighs for training
• Check if anything is missing from Athena Simclr