Hugging Face expands Open ASR Leaderboard with Egyptian and Nigerian English accents

If you’ve ever tried a voice assistant that stumbles over an Egyptian or Nigerian speaker, you’ve felt the bias that still exists in many speech‑recognition systems. The Open ASR Leaderboard, a community‑run benchmark hosted by Hugging Face, has just widened its scope by adding private test splits for those two accents, giving developers a concrete way to measure and improve performance on them.

The new splits are supplied by Liva AI, a company that gathers multimodal interaction data from real‑world consumer products. By embedding naturally occurring speech—complete with pauses, filler words and case‑sensitive punctuation—the data aims to reflect how people actually talk, rather than the clean, scripted recordings that dominate most public corpora today.

Source: Hugging Face: a]:hidden"> Egyptian and Nigerian English join the Open ASR Leaderboard ↗

Hugging Face announced the collaboration with Liva AI, stating that the two private splits contain 1.34 hours of Egyptian English and 1.55 hours of Nigerian English, both recorded in conversational, spontaneous settings. The recordings are punctuated, cased and include natural disfluencies, providing a realistic test bed for automatic‑speech‑recognition models.

Beyond the raw data, the leaderboard’s reporting format has been updated. Results are now broken out by “US vs. non‑US” groups and further divided into “Inner Circle” (US, UK, Australia, Canada) and “Outer & Expanding Circles” (India, Nigeria, Egypt). This mirrors Kachru’s Three‑Circles model and lets researchers see how well their systems serve speakers outside the traditional core English locales.

Source: Hugging Face: a]:hidden"> Egyptian and Nigerian English join the Open ASR Leaderboard ↗

The added splits are private, meaning they are not publicly downloadable and can only be evaluated through the leaderboard’s hosted inference API. Consequently, developers cannot directly fine‑tune models on the data, limiting the utility to benchmarking rather than training. Moreover, only two African accents are represented, leaving many other regional varieties still untested.

Because the benchmark still relies on a mix of private and public splits, performance gains on the new accents may not translate to real‑world deployments where acoustic conditions, background noise, and code‑switching are more varied. Until broader, openly available datasets emerge, the community will have to treat these results as a useful but partial indicator of global ASR fairness.

If you’re building or evaluating an ASR system, the quickest way to benefit is to submit your model to the Open ASR Leaderboard and see how it scores on the Egyptian and Nigerian splits. The leaderboard provides a free inference endpoint, so you can benchmark without needing the raw audio files.

You can also follow Hugging Face’s blog and the Liva AI project page for updates on new accents or languages that may be added later. Engaging with the community discussions around the “inner vs. outer circle” breakdown helps you understand where your model’s strengths and blind spots lie, guiding data‑collection priorities for more inclusive speech technology now.