Hugging Face has added two new private speech-recognition test sets to its Open ASR Leaderboard, covering Egyptian and Nigerian English. The data comes from Liva AI, which collected conversational, spontaneous speech for each accent. The Egyptian set runs to about 1.34 hours and the Nigerian set to about 1.55 hours. Both are punctuated, cased and include disfluencies such as false starts and filler words, so they reflect how people actually talk rather than read scripted text.
The additions build on earlier private data. Hugging Face previously added 11 splits from Appen and DataOceanAI covering scripted and conversational speech across five accents, and later an Indian-accented English split from Voice Arena. Private, held-out data matters because models can be tuned to public test sets, which inflates their scores. Hugging Face says the new sets are meant to widen accent coverage and reduce that risk by bringing in new data providers. The leaderboard will also now group results using Kachru's Three Circles model of English. It reports US versus non-US results, and Inner Circle (US, UK, Australia, Canada) versus Outer and Expanding Circle (India, Nigeria, Egypt) results. The team plans to split the Outer and Expanding Circles further as more accents are added. The breakdown idea came from Hume AI's leaderboard.
Liva AI gathers multimodal interaction data through consumer products that real people use, and it owns both the collection environment and the distribution channel. The company says this lets it target specific model weaknesses by domain, language and interaction type. For Australian businesses and professionals, this changes little directly. The new datasets are small and are used to test models, not to change products or contracts, and nothing here alters obligations under the Privacy Act. It is still worth checking the leaderboard's private-data results before choosing speech-to-text software for customer calls or transcription, particularly if your staff or customers speak English with accents that current benchmarks under-represent.
Key Points
Hugging Face added private Egyptian and Nigerian English speech test sets to the Open ASR Leaderboard, sourced from Liva AI.
The two sets total about 2.9 hours of conversational, spontaneous speech with disfluencies, punctuated and cased.
Private held-out data is intended to reduce benchmark overfitting and make leaderboard scores more trustworthy.
The leaderboard now breaks down accent results by US vs non-US and by Kachru's Inner Circle vs Outer and Expanding Circles.
Liva AI collects data through consumer products it owns and distributes, which it says captures natural, real-world speech.