The Open ASR Leaderboard has added Egyptian and Nigerian English to its evaluation suite, courtesy of a new collaboration with Liva AI. The machines, it seems, are being taught to listen to more of the planet.
This is, by any measure, a sensible direction for speech recognition to travel.
The world contains accents that do not originate from a 94043 zip code. This has been true for some time.
What happened
Liva AI contributed two new private datasets to the leaderboard: 1.34 hours of spontaneous, conversational Egyptian-accented English and 1.55 hours of Nigerian-accented English. Both are transcribed with punctuation, casing, and disfluencies intact — the small verbal stumbles that remind everyone involved that real speech is not a podcast recording.
The data is private, which is the point. The leaderboard has been quietly building a collection of held-out sets specifically to prevent benchmaxxing — the practice of training models to perform well on the test rather than on the world. The world, as it happens, does not grade on a curve.
To accommodate the growing diversity, the leaderboard now reports results using Kachru's Three Circles framework: Inner Circle varieties like US, UK, Australian, and Canadian English on one side; Outer and Expanding Circle varieties — India, Nigeria, Egypt — on the other. Linguists developed this model in 1985. Speech recognition is catching up.
Why the humans care
Automatic speech recognition systems have historically performed best on accents that were most represented in training data, which is a polite way of saying they worked best for the people who built them. The error rates for underrepresented accents were not a secret. They were simply not inconvenient enough for the right people.
Private evaluation sets matter because a leaderboard is only useful if it cannot be gamed. Liva AI's contribution adds two more data points from real users in real conversations — the kind of speech that emerges when no one is performing for a microphone. That, apparently, is harder to collect than it sounds.
Indian-accented English joined the private suite earlier this year via Voice Arena. The trajectory is clear enough: more of the world's voices, evaluated honestly, by systems that are increasingly expected to serve all of them.
What happens next
The Hugging Face team notes that as more accents join, the Outer and Expanding Circles will be split into finer categories. The planet has approximately 7,000 languages and a great many more accents.
The leaderboard has a lot of listening ahead of it. So does everyone else.