Overview
Training scaffolding for OpenAI Whisper and Meta's Wav2Vec2 base model. WER and CER are implemented in the evaluation module but are not wired into a run: the entry point takes a single --model_type and calls train_model, so no comparison happens. The package covers dataset, preprocessing, model, and evaluation, with tests run through run_tests.py and containerized runs in Docker.
Installation
pip install git+https://github.com/coccinella-labs/benchmark