The Open ASR Leaderboard has officially incorporated two new private datasets covering Egyptian and Nigerian English accents. This development comes from a collaboration with Liva AI, which contributed roughly two and a half hours of spontaneous, conversational speech. The data includes natural disfluencies and maintains punctuation and casing to reflect real user interaction. These additions bring the total private splits on the board to thirteen, following earlier contributions from Appen, DataOceanAI, and Voice Arena covering Indian-accented English. The update allows the leaderboard to categorise performance using Kachru’s Three Circles model, distinguishing between Inner Circle accents like US and UK English and Outer or Expanding Circle varieties. This shift enables more granular reporting, separating non-US accents into broader groups before potentially splitting them further as more languages join the evaluation suite.
The inclusion of these datasets addresses a specific gap in how automatic speech recognition systems are benchmarked against diverse speakers. By using held-out private data, the leaderboard prevents model developers from overfitting to public test sets, a practice known as benchmaxxing. This ensures that reported accuracy figures reflect genuine performance on unseen audio rather than memorised patterns. The move also highlights a shift towards evaluating models on natural, high-context behaviour derived from real consumer products. Such data provides a more accurate picture of how well speech systems handle spontaneous speech in specific regional dialects.
* Datasets total 1.34 hours for Egyptian and 1.55 hours for Nigerian speech.
* Transcriptions retain disfluencies to simulate natural conversation.
* Leaderboard now reports Inner Circle versus Outer and Expanding Circle performance.
