freewym/espresso

Built on PyTorch and fairseq, Espresso provides modular architectures including Transformer, Conformer, and Transducer models with support for both end-to-end and hybrid ASR training. It features distributed multi-GPU/node training, parallelized beam search decoding with integrated language model fusion, and on-the-fly feature extraction from raw waveforms via torchaudio. Recipes for WSJ, LibriSpeech, and Switchboard datasets demonstrate support for CTC, LF-MMI, and cross-entropy training objectives.

940 stars. No commits in the last 6 months.