espresso

Why we included this project

Espresso is aimed at speech researchers and engineers who want reproducible end-to-end ASR baselines without assembling the pieces themselves. It is a PyTorch-based training and evaluation toolkit that follows fairseq conventions, with YAML configs, distributed training across GPUs and nodes, and a choice of decoders that includes CTC, Transducer (RNN-T), and word-level language model fusion backed by a fast parallelized implementation. Ready-made recipes for WSJ, LibriSpeech, and Switchboard give newcomers a concrete place to start, and the optional torchaudio feature extraction path keeps Kaldi out of the pipeline for simpler setups.