Open ASR Leaderboard Adds Egyptian and Nigerian English

Speech recognition works best for the voices it was tested on, and that has historically meant a narrow slice of English accents. Hugging Face is widening the lens. On November 6, its Open ASR Leaderboard team announced a collaboration with Liva AI that adds two new held-out test sets for Egyptian and Nigerian English, plus a new way to compare how speech-to-text models perform across accent groups.

• Liva AI contributed conversational test sets for Egyptian English (1.34 hours) and Nigerian English (1.55 hours)
• Both sets use spontaneous speech, transcribed with punctuation, casing and disfluencies
• The leaderboard's Private tab now compares US versus non-US English accents
• It also groups accents using Kachru's Three Circles, separating Inner Circle from Outer and Expanding Circle speakers

Why Accent Coverage Matters for Speech Recognition

Automatic speech recognition (ASR) powers captions, voice assistants, meeting notes and accessibility tools. When a model stumbles on an accent, the people most affected are the very users who rely on those tools every day. Benchmarks shape what model builders optimize for, so adding underrepresented accents to a widely watched leaderboard nudges the whole field toward speech-to-text that works for more of the world.

The new sets are also conversational rather than read aloud. Spontaneous speech, with its restarts, fillers and overlapping thoughts, is much closer to how people actually talk to their devices, and it is a tougher, more honest test than scripted prompts.

How Do Private Test Sets Improve ASR Benchmarks?

The Egyptian and Nigerian splits join the leaderboard's private data, which earlier included 11 splits from Appen and DataOceanAI covering five accents, plus an Indian-accented English split added with Voice Arena. Because private data is held out from public release, models cannot be tuned against it, which Hugging Face says helps minimize "benchmaxxing" and keeps the rankings trustworthy.

What the New Accent Breakdown Shows

With the new data, the leaderboard reports two comparisons. The first is simple: US English versus non-US English, covering the UK, Australia, Canada, India, Nigeria and Egypt. The second borrows Kachru's Three Circles model of English use, grouping the US, UK, Australia and Canada as the Inner Circle, and India, Nigeria and Egypt as the Outer and Expanding Circles. Hugging Face credits Hume AI with the idea and says it will split the outer groups further as more accents join.

For model builders, that view makes accent gaps visible at a glance. For buyers choosing a transcription model, it is a practical way to check whether a top-ranked model also serves your actual users.

If you want to understand how these scores are calculated and where they mislead, read our explainer on speech recognition benchmark diagnostics, and see a strong open contender in our look at IBM Granite Speech 5. More model news lives in our AI section.

Sources: Hugging Face blog: Egyptian and Nigerian English join the Open ASR Leaderboard — November 6, 2026.