The Open ASR Leaderboard has added private test sets for Egyptian and Nigerian English, expanding how it measures speech-recognition systems beyond accents commonly represented in public benchmarks. The data comes through a collaboration with Liva AI.
Because the audio is held back from model developers, it can provide a stronger check against “benchmark maxing,” where a system is tuned to known evaluation material without improving more broadly. The leaderboard already includes private splits covering accents from the United States, United Kingdom, Australia, Canada and India.
Results will now include comparisons between US and non-US speech, as well as broader groupings based on linguist Braj Kachru’s model of global English use. Hugging Face says it plans to separate those categories further as more accents become available.
Accent coverage matters because an average score can hide large recognition gaps for particular communities. Two additional test sets do not make the benchmark globally representative, but they give developers a clearer view of whether gains transfer to English spoken in Egypt and Nigeria rather than only to familiar public datasets.