openspeech

Why we included this project

OpenSpeech is a training toolkit for end-to-end automatic speech recognition, built around working reproductions of published ASR models such as DeepSpeech2 and RNN-Transducer. It includes training recipes for English, Chinese, and Korean, with datasets like LibriSpeech, AISHELL-1, and KsponSpeech, so you can reproduce a paper's results or point it at your own data by defining a data module and tokenizer. Since the code runs on PyTorch-Lightning and Hydra, mixed-precision, multi-GPU and TPU training, and hierarchical config management are already wired in, which spares you most of the plumbing that slows down ASR research. One caveat: this is a training framework, not a drop-in inference service, so plan to train or fine-tune models yourself rather than expecting pretrained weights.