Hugging Face announced on November 6, 2026 that its Open ASR Leaderboard now includes Egyptian- and Nigerian-accented English. For customer-service teams, businesses and creators choosing transcription software, the addition broadens the evidence available about the voices in their recordings.
It is an expansion of testing, not an upgrade to a transcription product. It does not establish that any system now understands those speakers better—or identify the best tool for either accent.
Liva AI has supplied private conversational recordings covering Egyptian and Nigerian English: 1.34 hours and 1.55 hours respectively, according to the announcement. These are spontaneous conversations in English, not new tests of Arabic or Nigeria’s other languages.
That is approximately 2.89 hours combined, calculated by adding the listed durations. The dataset table does not supply speaker counts for the two sets and lists their sex distribution as unavailable. Neither the duration nor a country label establishes coverage of every regional variety, age group or recording environment.
A group average is not an accent-specific verdict
The leaderboard’s Private data comparisons include US versus non-US accents, and a view combining India, Nigeria and Egypt opposite the US, UK, Australia and Canada. Individual private test-set scores are withheld. A strong result for a mixed group therefore cannot establish a strong result for each accent within it.
The weighting also matters. The current scoring code combines three Indian datasets with one Nigerian and one Egyptian dataset in the India–Nigeria–Egypt group. Assuming all five have usable positive scores, the Indian datasets supply 3/5, or 60%, of that average; Nigeria and Egypt each supply 1/5, or 20%.
Those percentages are a calculation of dataset weights, not accuracy results or shares of speakers. Missing scores can change the weights. Even with complete coverage, a model could lead the combined group without leading on the recordings most relevant to a particular buyer.
The November 6 configuration includes private scripted and conversational averages in the main leaderboard’s default dataset selection. For a fair comparison over time, keep the leaderboard version and selected datasets alongside the result: a changed test mix is not evidence that the model itself improved.
What an error score leaves out
The main accuracy measure is word error rate, or WER: substituted, omitted and added words divided by the number of words in the reference transcript. Lower is better.
As an illustrative calculation, four substitutions, three omissions and three additions against a 100-word reference give (4 + 3 + 3) / 100 = 10% WER. That does not mean 90% of customer calls were handled correctly. A wrong customer name and an inconsequential word each contribute an error, although their business consequences may differ.
The benchmark also normalizes text before scoring, removing punctuation, capitalization and filler words, among other adjustments. Its WER score therefore does not establish whether a transcript has readable punctuation or publication-ready captions.
Private recordings involve a tradeoff
Hugging Face says withholding test audio and individual-set scores is intended to reduce optimization aimed narrowly at the benchmark. Its stated process has maintainers check submitted models’ public results and run the private evaluations.
The tradeoff is that an outside buyer cannot directly reproduce those private scores without the recordings and reference transcripts. Public evaluation code helps explain the calculation; it does not make the underlying material independently inspectable. Keeping test data private is a design choice, not proof that a score will transfer to a buyer’s own recordings.
Ask for evidence that matches the job
The wider research supports looking beyond a universal ranking. The authors of the 2025 AfriSpeech-MultiBench study, covering more than 100 African English accents, reported variation by country and application. That is separate evidence for task-specific evaluation, not validation of Liva’s new recordings.
Use the Open ASR Leaderboard to build a shortlist, then seek a small, representative evaluation before choosing a service:
Match the recordings to the work. Use audio you have permission to use, with the speakers, devices, background noise and vocabulary the service will encounter. Compare the same recordings against independently checked transcripts.
Check what mistakes cost. Alongside word errors, examine names, amounts, missing details and correction time. If the task needs speaker labels or readable captions, assess those separately.
Compare the service, not just the model. Ask which model version and settings produced the evidence, and check current price, processing delay and data-handling terms. The accent announcement does not settle those purchasing questions.
The same distinction between a headline capability and the actual task appears in RohitAI’s explanation of document-tracking benchmarks. Here, the useful advance is broader accent coverage in evaluation. The remaining buying question is how well a shortlisted tool transcribes the people and details that matter to your work.