Condenser microphone hanging over a mixing console in a recording studio

· Translated from the Spanish original · 中文 · 日本語

Speech recognition

The Open ASR table adds private data on Egyptian and Nigerian accents to measure better

Hugging Face and Liva AI add two English accents with unpublished data, to curb overfitting and show performance by accent groups.

What happened

  • Hugging Face announced on 6 November 2026 a collaboration with Liva AI to add private data from two new accents, Egyptian and Nigerian, to the Open ASR Leaderboard, its automatic speech recognition table.
  • A few months ago the table had incorporated private data from Appen and DataOceanAI: 11 splits of scripted and conversational speech from five accents. It then added one more, Indian-accented English, with Voice Arena.
  • With these new accents, the table now shows a richer breakdown, based on Kachru’s Three Circles: the Inner Circle (United States, United Kingdom, Australia, Canada) against the Outer and Expanding Circles (India, Nigeria, Egypt).
  • The idea of the breakdown, Hugging Face acknowledges, comes from Hume AI’s leaderboard.

Why it matters

  • A model that understands American English speakers well may perform worse with other accents, and an overall average hides that difference.
  • Private data is not published, so models cannot be trained on it. That reduces what Hugging Face calls “benchmaxxing”, that is, optimizing for the test rather than for real use.
  • More accents and more languages in evaluations improve representation of those who use these systems outside English-speaking countries.

The number

Five, plus one, plus two. The table started with 11 private splits covering five accents, later added one of Indian-accented English and now adds two more accents: Egyptian and Nigerian.

Context

Liva AI collects multimodal interaction data through consumer products used by real people. According to Hugging Face, because it controls the collection environment and the distribution, it can design products around specific model gaps, by domain, language or modality, and obtain natural, high-context behaviour rather than elicited tasks.

That description is Hugging Face’s and reflects the partners’ point of view. The text gives no model results on the new accents and no count of the hours of audio added.

The ASR table works alongside another, the Open TTS Leaderboard, for text-to-speech models. Both aim to give developers a common place to compare open models.

At Mamífero we follow speech recognition and synthesis. We reported that Nvidia released Nemotron 3.5 Transcribe as an open model and that Phonon generates speech offline in 75 MB. Also that Impression-R1 drafts radiology conclusions and that rewarding each search step makes an agent more reliable.

What’s next

  • Speech recognition models will now also be evaluated by accent groups.
  • Hugging Face says it will keep adding accents, languages and data providers.
  • Model results on these new private sets have not yet been published in the announcement.

Bottom line

A public test can be studied until it is memorized. With data nobody sees and more accents, the table tries to measure whether a model truly understands the person speaking to it.

Sources

  • Hugging Face, “Egyptian and Nigerian English join the Open ASR Leaderboard”, 6 November 2026, link.
  • Cover image: Josephenus P. Riley, CC BY 2.0, via Wikimedia Commons

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes