Egyptian and Nigerian English join the Open ASR Leaderboard
Reported by Hugging Face and ecosistemastartup2 sources ↓
Hugging Face says it has added private Egyptian- and Nigerian-accented English data to the Open ASR Leaderboard through a collaboration with Liva AI.
Key points
- The new data is private and intended to reduce optimization against publicly known evaluation sets
- The two datasets contain spontaneous conversational speech, rather than scripted recordings
- Hugging Face will report accent breakdowns using Kachru’s Three Circles framework
Sources
Details
The datasets were added through a collaboration between Hugging Face and Liva AI. Ecosistemastartup reports that the Egyptian-accented English set contains 1.34 hours of audio and the Nigerian-accented English set 1.55 hours; both are conversational recordings transcribed with punctuation, capitalization and disfluencies.
Hugging Face says private evaluation data can help limit “benchmaxxing,” or tuning models to perform well on known public test sets. The leaderboard had previously added private speech data from Appen and DataOceanAI, and a private Indian-accented English split with Voice Arena.
The new accents also allow the leaderboard to provide a richer breakdown based on Kachru’s Three Circles of English. Hugging Face says it plans to separate the Outer and Expanding Circles as more accents are added.
Written from coverage as of
Images
- Hugging Face
- ecosistemastartup
- Hugging Face
- Hugging Face