African Language Speech Recognition: Why Hausa, Yoruba and Amharic Still Lag
24
Around 94 million people speak Hausa. Around 60 million speak Amharic, and about 50 million speak Yoruba [1][2][3]. These are not small languages. Yet African language speech recognition still gets between a quarter and half of the words wrong for them when the audio comes from real life rather than a recording booth. If you are building a voice product for Africa, or evaluating one, that gap decides whether it works.
What the latest benchmark shows
AfriVox-v2, an in-the-wild African speech benchmark published in May 2026, tests current systems on audio recorded in real environments rather than read in quiet rooms [4]. The word error rates (WER) on its in-the-wild split are sobering:
Language | Speakers (approx.) | WER range across tested systems |
|---|---|---|
Hausa | ~94M | 26.9% to 50.2% |
Amharic | ~60M | 24.9% to 49.0% |
Yoruba | ~50M | 29.6% to 43.8% |
Igbo | ~34M | 28.7% to 45.9% |
Speaker counts: [1][2][3][7]. Read the bottom of each range carefully: even the best system tested gets roughly one word in four wrong in Hausa and Amharic. General-purpose systems from the major labs sit in the 25 to 50% band. For a voice agent, a transcription product or a call-centre analytics tool, that is the difference between a usable product and one that frustrates users in their own language.
Not every African language is “low-resource”
It is tempting to file all of this under “low-resource languages” and move on. The same benchmark shows why that label hides more than it explains. On Swahili, the best systems tested on AfriVox-v2 score around 7% WER [4]. That is close to what many teams accept for production English.
Swahili is not structurally easier than Hausa or Yoruba. It has simply had more attention: more public data, more evaluation and more commercial demand. The difference between 7% and 27 to 50% is mostly a difference in data, and that is good news, because data is something you can fix.
Why the gap persists
Three things keep error rates high for languages like Hausa, Yoruba and Amharic.
Too little real-world audio. Public corpora for these languages are small, and much of what exists is read speech, broadcast audio or scraped content with uncertain labels. Models trained on that do not hear how people actually talk on their own phones, in markets, homes and offices.
Transcripts that lose the language. Yoruba is a tonal language whose meaning depends on tone marks. Amharic is written in Ge’ez script. Transcripts that strip diacritics or romanise the script teach a model a simplified version of the language and make evaluation unreliable.
Accents, dialects and code-switching. Speakers mix languages mid-sentence and vary by region. A test set from one city or one demographic overstates quality everywhere else.
Targeted, verified data closes the gap
The encouraging part is how much targeted data moves the numbers. A September 2026 study on Minnan (Hokkien) dialects fine-tuned an open-source multilingual ASR model on a dedicated dialect corpus built with human-verified annotation. Its character error rate fell from 34.6% to 17.6%, roughly halving the error [5].
The lesson carries across languages: when the training and evaluation data match the speakers you serve, in their script, accent and environment, accuracy improves sharply. The Swahili result shows where that path leads.
What to ask before you launch in an African language
If you are planning African language speech recognition for a product, ask these questions of any model or data supplier:
Was the test set recorded in the wild? Read speech in quiet rooms will flatter every system.
Is the transcript in the native script, with tone marks where the language needs them?
Is the speaker metadata verified? First language, country, accent, age band and device should travel with every clip.
Is there a consent record for every speaker? Buyers and regulators increasingly expect proof, not promises.
Can you get more of the same? A benchmark is a starting point; you will need training data that matches it.
Where Silencio fits
Silencio gives AI teams access to a pipeline of about 2.5 million diverse speakers across roughly 350 languages and 180 countries. Contributors record on their own devices in real environments, and every recording carries verified speaker metadata and a consent record for every speaker.
For the languages in this article, our catalogue includes more than 10,000 hours of human-verified Amharic in Ge’ez script, more than 10,000 hours of human-verified, tone-marked Yoruba, and 5,000 to 10,000 hours of human-verified Hausa. Where a language, dialect or accent is not in the catalogue, we collect it to spec with native speakers where it is spoken.
Public samples, including native-script Yoruba, Hausa and Amharic, are on Hugging Face [6].
FAQ
What is a good word error rate for speech recognition?
It depends on the use case, but many production English systems run well under 10% WER on clean audio. In-the-wild results for Hausa, Yoruba and Amharic are currently 25 to 50%.
Why is Swahili ASR so much better than Hausa or Yoruba ASR?
Mostly data and attention. Swahili has had more public data, more benchmarks and more commercial demand. Targeted data can close much of the gap for other languages.
Do I need native-script transcripts?
Yes. Romanised or diacritic-free transcripts lose meaning in tonal languages such as Yoruba and make evaluation unreliable.
How much data do I need to improve a language?
It varies by language and starting point. Published results show that a dedicated dataset for the target language or dialect can roughly halve the error rate of an existing multilingual model, and a small in-the-wild evaluation set is the right first step to size the gap.
Want an evaluation set or training data for an African language on your roadmap? Talk to us at silencio.network.
Sources
Hausa language, Wikipedia (citing Ethnologue). https://en.wikipedia.org/wiki/Hausa_language
Amharic, Wikipedia (citing Ethnologue). https://en.wikipedia.org/wiki/Amharic
Yoruba language, Wikipedia (citing Ethnologue). https://en.wikipedia.org/wiki/Yoruba_language
Awobade, Ashungafac and Olatunji, “AfriVox-v2: A Domain-Verticalized Benchmark for In-the-Wild African Speech Recognition”, arXiv 2605.03590, May 2026, Table 5 (in-the-wild split). https://arxiv.org/abs/2605.03590
“WenetSpeech-Min”, arXiv 2609.36834, September 2026, Table 2. https://arxiv.org/abs/2609.36834
Silencio on Hugging Face. https://huggingface.co/SilencioNetwork
Igbo language, Wikipedia. https://en.wikipedia.org/wiki/Igbo_language