neural_sp

Why we included this project

For anyone building speech recognition systems, neural_sp is a handy PyTorch reference implementation that pulls together the main ASR building blocks in one place. You get the full pipeline: feature front-ends like SpecAugment, encoders ranging from RNNs to Transformers and Conformers, and decoders for CTC, RNN-Transducer, and attention-based approaches, so comparing architectures doesn't mean gluing together several libraries. Streaming is a real strength here, with monotonic attention, MoChA, and latency-controlled variants implemented, which matters if low-latency recognition is your goal rather than offline batch decoding. The repo also includes recipes for common corpora like LibriSpeech, WSJ, Switchboard, and AISHELL, along with published word and character error rates you can reproduce as baselines. It's a compact testbed for evaluating which encoder or decoder.