Open ASR Leaderboard Adds Egyptian and Nigerian English Tests
Open ASR Leaderboard adds 1.34 hours of Egyptian and 1.55 hours of Nigerian English to strengthen private, accent-aware speech recognition evaluations.
Summary
On November 6, 2026, the Open ASR Leaderboard added two private Liva AI datasets: 1.34 hours of Egyptian English and 1.55 hours of Nigerian English. Both contain spontaneous conversational speech transcribed with punctuation, casing and disfluencies. They follow 11 private Appen and DataOceanAI splits covering scripted and conversational speech across five accents, plus a Voice Arena split for Indian-accented English. Held-out data is intended to reduce benchmaxxing, strengthen evaluation trust and broaden accent representation.
The Private tab now compares US speech with non-US speech from the UK, Australia, Canada, India, Nigeria and Egypt. Using Kachru’s Three Circles, an approach credited to Hume AI, it also compares the Inner Circle, comprising the US, UK, Australia and Canada, with the combined Outer and Expanding Circles of India, Nigeria and Egypt; the latter groups will separate as more accents arrive. Liva AI gathers natural, high-context multimodal interactions through consumer products it owns and distributes, enabling data collection around model gaps across domains, industries, languages, modalities and interaction types.
Positives
- 1.34 hours of Egyptian English add a previously unrepresented accent to the leaderboard’s private evaluations.
- 1.55 hours of Nigerian English expand coverage of Global South English speech.
- Both Liva AI datasets capture spontaneous conversation with punctuation, casing and disfluencies.
- Private held-out datasets are intended to reduce benchmaxxing and improve confidence in ASR comparisons.
- Seven accents now support US versus non-US and Kachru-based performance breakdowns in the Private tab.
Risks & concerns
- The Egyptian and Nigerian datasets contain just 1.34 and 1.55 hours of speech, respectively.
- Both additions cover spontaneous conversational English, adding neither scripted speech nor another language in this update.
- Outer and Expanding Circle accents remain combined until the leaderboard gains enough coverage to separate them.