Leaderboard

Why we included this project

Speech recognition is only as trustworthy as the data you test it on, and this project makes that testing unusually easy. It bundles a large set of test corpora in English and Chinese, from academic standards like LibriSpeech and AISHELL to messier real-world material such as live broadcasts, podcasts, and TV interviews, each tagged with a difficulty rating. Alongside the data sits a model zoo of commercial APIs and open-source recognizers, plus a pipeline that handles data preparation, recognition, post-processing, and error-rate scoring. The result is that you can reproduce published results, run your own model against a wide range of systems on identical data, and see how a recognizer holds up outside clean academic conditions. For anyone choosing or tuning an ASR engine, the professionally transcribed test sets are a good way to surface weaknesses that standard benchmarks miss.