Hugging Face Adds Egyptian and Nigerian Dialects to Open ASR Leaderboard
By Ali Al-Montaser — Senior AI Technical Editor
Hugging Face, in partnership with Liva AI, has added Egyptian (1.34 hours) and Nigerian (1.55 hours) dialects to the Open ASR Leaderboard, totaling 2.89 hours of conversational speech data as private splits to reduce benchmaxxing. The update restructures dialect analysis using Kachru's Three Circles and is available immediately on Hugging Face Spaces.
Executive Overview
Hugging Face has officially announced a new collaboration with Liva AI to add two new dialects to the Open ASR Leaderboard: Egyptian (1.34 hours) and Nigerian (1.55 hours) of conversational spontaneous speech, totaling 2.89 hours of private split data. The update is available immediately via the Hugging Face platform and is part of efforts to improve dialect representation and reduce benchmaxxing in the evaluation of automatic speech recognition systems.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Access to the Open ASR Leaderboard is free and open via the Hugging Face platform. There are no fees for evaluation or viewing results. Liva AI's private data is not sold as a separate product; it is used for evaluation purposes only. |
| 🌐 Platforms & Immediate Availability | Available immediately via Hugging Face Spaces (two Spaces mentioned in the article). Access via the web directly, with the ability to programmatically call results via the Hugging Face Hub API. |
| ⚡ Performance & Speed Benchmarks | No WER (Word Error Rate) figures are published in the original announcement for this addition; the update focuses on adding data and restructuring the analysis. Confirmed quantitative metrics: 1.34 hours (Egypt) + 1.55 hours (Nigeria) = 2.89 total hours. |
| 🛡️ Security & Anti-Cheat Resistance | Data is private (Private Split) and held-out to reduce benchmaxxing. Models are not exposed to test data, limiting benchmark manipulation. No explicit security standards are mentioned otherwise. |
| 🧠 Context Window | Not applicable — this is an ASR evaluation leaderboard, not a language model. Audio data size is the metric: 1.34 hours and 1.55 hours. |
| 🌍 Arabic Language & Regional Support | The Egyptian dialect (EG) is now officially represented in the private evaluation. Texts are punctuated, cased, and contain disfluencies, reflecting natural Arabic speech. The Nigerian dialect is Nigerian English (NG). |
Deep-Dive Features & Architecture
This update continues a path Hugging Face began months ago by adding 11 private splits from Appen and DataOceanAI covering five dialects of read and conversational speech, followed by a private split for the Indian dialect in collaboration with Voice Arena. The new addition with Liva AI brings the total to seven dialects represented in the private data. The new data is classified as Conversational, spontaneous with Punctuated, cased, with disfluencies text, reflecting natural unprepared speech—an important indicator for ASR systems dealing with real conversations.
Liva AI collects data through consumer products used by real people, giving it control over both the collection and distribution environment, and allowing it to design products around specific model gaps across domains, industries, languages, modalities, and interaction types. This approach yields natural, high-context behavioral data rather than artificially elicited or task-based data, enhancing evaluation reliability.
On the analysis structure front, Hugging Face adopted the Kachru’s Three Circles classification (thanks to Hume AI for using it in their leaderboard), making the report two-tiered: first, American vs. non-American (UK, Australia, Canada, India, Nigeria, Egypt), and second, Inner Circle (US, UK, Australia, Canada) vs. Outer and Expanding Circles (India, Nigeria, Egypt). Hugging Face noted a plan to split the Outer and Expanding Circles as new dialects are added, meaning the structure is scalable.
Benchmark & Competitive Performance
The original announcement does not include WER figures or direct numerical comparisons between models on the two new dialects, but the addition strengthens the comparison infrastructure by providing held-out sets that reduce benchmaxxing—i.e., optimizing models on public test data. In this sense, the competitive value lies in the reliability of the benchmark rather than in specific performance numbers. The private splits ensure that models are evaluated on unseen data, promoting fairer comparisons and encouraging generalization rather than overfitting to public benchmarks.
Industry Impact & Enterprise Adoption
This update has significant implications for enterprises deploying ASR systems in multilingual and dialect-rich environments. By including Egyptian and Nigerian dialects, the Open ASR Leaderboard now offers a more realistic assessment of ASR performance in regions where dialectal variation is common. Enterprises can use these benchmarks to select models that perform well on their target dialects, reducing deployment risks. The private split approach also sets a precedent for other benchmarks to follow, emphasizing data integrity and fair evaluation. As Hugging Face continues to expand dialect coverage, the leaderboard becomes an increasingly valuable resource for developers and organizations aiming to build inclusive speech technologies.
Conclusion
The addition of Egyptian and Nigerian dialects to the Open ASR Leaderboard marks a step forward in dialect representation and evaluation integrity. With 2.89 hours of private conversational data and a restructured analysis based on Kachru’s Three Circles, Hugging Face and Liva AI are addressing the critical need for robust, unbiased ASR benchmarks. While specific WER numbers are not yet available, the infrastructure improvements promise more reliable comparisons and encourage the development of ASR systems that generalize across dialects. This update is available immediately and free of charge, reinforcing Hugging Face’s commitment to open and accessible AI evaluation.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Hugging Face, through Liva AI, added 1.34 hours for the Egyptian dialect and 1.55 hours for the Nigerian dialect, totaling 2.89 hours of conversational, spontaneous speech with punctuated, cased text containing disfluencies.
The data provider is Liva AI, which collects multimodal interaction data through consumer products used by real people, enabling natural, high-context behavioral data rather than synthetic task-based data.
The update adopts Kachru's Three Circles: an initial classification between American and non-American (UK, Australia, Canada, India, Nigeria, Egypt), then a second classification between the Inner Circle (US, UK, Australia, Canada) and the Outer and Expanding Circles (India, Nigeria, Egypt), with plans to split the Outer and Expanding Circles as new dialects are added.
The data is private (Private Split) and is added to the Private tab in the Open ASR Leaderboard, aiming to reduce benchmaxxing and increase evaluation reliability through held-out sets.
Hugging Face previously added 11 splits from Appen and DataOceanAI covering five dialects of read and conversational speech, then added a private split for the Indian dialect in collaboration with Voice Arena, bringing the total to seven dialects represented in the private data with Egypt and Nigeria.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
