SMAR Daily Industry News

Open ASR Leaderboard Adds Egyptian and Nigerian English

The Open ASR Leaderboard has expanded its speech-recognition model evaluations by incorporating two new accents: Egyptian and Nigerian English. This update is the result of a new collaboration with Liva AI, a company that gathers multimodal interaction data through consumer products.

The addition includes two specific private data splits. The Egyptian dataset provides 1.34 hours of conversational, spontaneous speech, while the Nigerian dataset offers 1.55 hours of similar content. Both datasets feature punctuated and cased transcriptions that include disfluencies, reflecting natural speech patterns.

Liva AI collects this data through real-world consumer product usage rather than task-based or artificially elicited interactions. By controlling both the collection environment and distribution, the company aims to address model gaps across various domains, languages, and modalities. This expansion follows previous efforts to include private data from Appen, DataOceanAI, and Voice Arena, which covered five other accents including Indian-accented English.

The Open ASR Leaderboard is also updating its reporting structure to include a breakdown based on Kachru’s Three Circles model. As more accents are added, the platform plans to further segment the Outer and Expanding Circles to provide more granular performance insights.

For industry professionals, these additions represent a shift toward more inclusive and representative AI benchmarks, which is essential for reducing bias and improving the reliability of speech-recognition systems in global markets.

Why it matters

Expanding benchmark datasets to include diverse accents is critical for developers and businesses aiming to deploy globally accessible and accurate voice-AI technology.

FAQ

What new accents were added to the Open ASR Leaderboard?
The leaderboard added Egyptian and Nigerian English accents.

How does Liva AI collect its data?
Liva AI collects multimodal interaction data through consumer products used by real people, rather than using task-based or artificially elicited interactions.

What is the total duration of the new datasets?
The Egyptian dataset contains 1.34 hours of speech, and the Nigerian dataset contains 1.55 hours.

Source: Hugging Face Blog. See the original source for full details.