The Open ASR Leaderboard has added private evaluation speech with Egyptian and Nigerian accents through a collaboration with Liva AI. Both additions contain spontaneous conversation. The expansion also brings more detailed accent groupings, extending a testing programme that already included private data from other providers and an Indian English split.
The Open ASR Leaderboard, a comparison platform for automatic speech recognition, has added private Egyptian-accented and Nigerian-accented English evaluation splits with Liva AI, an interaction-data company. Both use spontaneous conversation rather than a prepared text read aloud. The Egyptian split contains 1.34 hours of audio and the Nigerian split 1.55 hours. Reference transcripts preserve punctuation, letter casing and disfluencies in the speech.[1], [2]
The private-data tab now compares US English with a group covering the UK, Australia, Canada, India, Nigeria and Egypt. A second breakdown uses Kachru’s Three Circles, an approach to grouping English use: the US, UK, Australia and Canada form the Inner Circle, while India, Nigeria and Egypt appear together in the Outer and Expanding Circles. The authors plan to separate those latter circles as more accents join.[1]
The additions follow 11 private splits supplied by Appen and DataOceanAI, evaluation-data providers, across five accents. Voice Arena, a voice-evaluation collaborator, later supplied an Indian English split. The authors describe new private data as a way to reduce tuning specifically to known tests and to broaden representation. Liva AI says it collects interaction data through consumer products used by real people.[1]
