wer_are_we

Why we included this project

Word-error-rate numbers are scattered across dozens of papers, and this repository gathers them into one place, benchmark by benchmark. On LibriSpeech alone you can compare the latest self-supervised models against older hybrid systems, with each row linking to the paper and noting the architecture and training data behind the score. It is a hand-maintained bibliography rather than deployable software, and the open invitation to correct the tables keeps the entries grounded in what people actually reported. If you are benchmarking a new acoustic model, writing a survey, or just trying to pick a sensible baseline, this is a fast way to find the primary papers worth reading in depth.