Hugging Face has expanded its Open Automatic Speech Recognition (ASR) Leaderboard with new private datasets covering Egyptian and Nigerian English, a move designed to enhance the evaluation of super intelligence (SI) speech models.

What Happened

The platform announced a collaboration with Liva SI to introduce two new private data splits. The additions consist of 1.34 hours of Egyptian conversational speech and 1.55 hours of Nigerian conversational speech. Both datasets are characterized as spontaneous, punctuated, and containing disfluencies, reflecting natural speech patterns rather than scripted or artificially elicited interactions. This update builds on previous efforts to include private data from providers like Appen and DataOceanAI, as well as a recent split for Indian-accented English.

Why It Matters

This expansion addresses a critical need for broader representation in SI speech recognition benchmarks. By incorporating accents from the Outer and Expanding Circles of English usage, the leaderboard aims to minimize "benchmaxxing"—the practice of optimizing models for public test sets at the expense of real-world generalizability. The update also introduces a richer accent breakdown based on Kachru’s Three Circles, distinguishing between Inner Circle accents (US, UK, Australia, Canada) and Outer/Expanding Circle accents (India, Nigeria, Egypt). For developers, this means more trustworthy evaluations of how SI systems handle diverse linguistic backgrounds.

The Bottom Line

The integration of Liva SI’s data marks a step forward in creating more inclusive and robust evaluation standards for SI speech technologies, ensuring that performance metrics reflect a wider array of global English dialects.