syhw/wer_are_we

Comprehensive leaderboards comparing Word Error Rate (WER) across standardized benchmarks (LibriSpeech, WSJ) with linked papers, architectures, and training methodologies. The project catalogues end-to-end and hybrid ASR approaches—from HMM-DNNs to Conformers and self-supervised models—documenting specific techniques like SpecAugment, language model rescoring, and data augmentation strategies. Community-maintained tables enable quick identification of SOTA results by dataset and publication date, serving as a bibliography for tracking architectural innovations and performance trajectories in speech recognition.

1,865 stars. No commits in the last 6 months.