Hugging Face Open ASR Leaderboard adds Egyptian and Nigerian English datasets
Overview of New ASR Datasets
Hugging Face has integrated two new private datasets into the Open ASR Leaderboard, specifically targeting Egyptian and Nigerian English accents. Provided by Liva AI, these datasets consist of conversational, spontaneous speech, which helps address the need for more diverse and representative evaluation metrics in automatic speech recognition (ASR) systems.
Dataset Specifications
The new additions include the following data:
| Dataset | Accent | Duration | Style | Transcription Details |
|---|---|---|---|---|
| Liva AI Conversational EG | Egyptian | 1.34 hours | Conversational, spontaneous | Punctuated, cased, with disfluencies |
| Liva AI Conversational NG | Nigerian | 1.55 hours | Conversational, spontaneous | Punctuated, cased, with disfluencies |
Enhanced Accent Categorization
To provide more granular performance insights, the Open ASR Leaderboard has updated its reporting structure based on Kachru’s Three Circles of English. The leaderboard now categorizes performance across two new dimensions:
- US vs. non-US: Includes UK, Australia, Canada, India, Nigeria, and Egypt.
- Inner Circle vs. Outer & Expanding Circles: Compares Inner Circle nations (US, UK, Australia, Canada) against Outer and Expanding Circle nations (India, Nigeria, Egypt).
Hugging Face intends to further segment the Outer and Expanding Circle categories as more regional accents are added to the platform.
Data Provenance and Methodology
The inclusion of this data is part of a broader initiative to minimize "benchmaxxing"—the practice of optimizing models specifically for known public test sets—by utilizing held-out, private data. Liva AI, the data provider, collects this information through multimodal interactions within consumer products. This approach allows for the capture of natural, high-context speech behavior that occurs in real-world environments, rather than relying on artificially elicited or task-based speech samples.