Research
4
benchmarks, telephony-channel and entity-dense
Open
benchmarks and cultural corpus, published freely
CC BY
licence on everything the public corpus holds
Every serious claim in this market is currently unverifiable, including the favourable ones we would like to make. The fastest route to credibility is to publish the evaluation rather than the result — an open, reproducible, telephony-realistic benchmark that anybody can run against anybody’s system, ours included.
Benchmarks
Nobody benchmarks 8 kHz African telephony audio with local surnames, regional currencies and code-switched dialects in it. So nobody optimises for it, so everybody is bad at it — and a buyer with exactly that problem has nowhere to go for a number they can trust.
8 kHz narrowband, code-switched, entity-dense. Every existing African speech benchmark measures clean audio, which is why every system scores well on them and badly on a phone call.
Key regional languages · 2,400 utterances · named entities and local currencies labelled
Person names, bank and fintech names, place names and currency, scored separately from the surrounding speech. Frontier systems run 40–70% WER here against 25–35% overall, and no public benchmark isolates it.
4,000+ entities · Surface forms across high-traffic African language clusters
Domain-matched rather than scripture-matched. A large share of available African-language text is religious, news or government notice, and none of it resembles a support conversation.
Financial services, healthcare and civic domains
Dialect identification within a language, which is where collection cost and model behaviour actually diverge — and which almost nothing measures.
Dialect clusters across East, West, and Southern African languages
The gap
The gap
This is not our benchmark. It is what the incumbents publish about themselves — word error rate bands, language lists and model tiers, read off their own documentation. The gap is documented rather than inferred, which is what makes it a business rather than an opinion.
| Vendor | Text to speech | Speech to text | What is missing |
|---|---|---|---|
| ElevenLabs | Text to speech6 African languages in v3 | Speech to textNo African language in its top band | What is missingYoruba, Igbo, Zulu and Amharic absent from TTS entirely |
| Text to speechAfrikaans only on Chirp 3 HD | Speech to textHausa, Yoruba, Igbo and Zulu capped at Chirp / Chirp 2 | What is missingAmharic, Afrikaans and Arabic get the current model; West African languages do not | |
| Microsoft Azure | Text to speechUnverified | Speech to text6 African languages | What is missingNeglects key indigenous African languages — English accent layers are partially served, indigenous language layers are not |
| AWS | Text to speechNone | Speech to textHausa, batch only | What is missingYoruba and Igbo missing from Transcribe entirely; Polly ships zero African voices |
ElevenLabs’ own published accuracy bands
Excellent · ≤5% WER
None
High · 5–10%
Swahili
Good · 10–20%
Afrikaans, Hausa, Arabic
Moderate · 25–50%
Yoruba, Igbo, Zulu, Xhosa, Amharic, Somali, Wolof, Shona, N. Sotho, Ganda
Swahili is the only African language above the Moderate band — which is also why it is the least attractive large African language to build for commercially. The opportunity is where the bands are worst.
Ships first
Ranked on commercial pull minus competitive crowding across African enterprise corridors. Built for the foundational lingua francas and high-volume commercial dialects that power real customer support, cross-border trade, and mobile banking across the continent.
The lingua franca of urban informal commerce from Lagos to Accra to Douala, supported by no major vendor at all, with no tonal orthography problem. The single best first language for a West African voice product.
35 AI readiness
80m±
Where the enterprise volume actually sits. Frontier ASR runs 24.7–35% WER on the accent generally — and 40–70% on African named entities, which is the number that decides whether a bank can deploy.
76 AI readiness
110m
Lagos-concentrated, so co-located with the buyers. Absent from ElevenLabs TTS, AWS Transcribe and Google Chirp 3; ElevenLabs' own published STT band for it is 25–50% WER.
49 AI readiness
46m
The best cross-border expansion asset in West Africa — one language investment reaches Niger, Chad, Ghana and northern Cameroon. Also the one West African language ElevenLabs does cover, so the deficit here is smaller than it looks.
65 AI readiness
88m
Commercially dense in trade and diaspora remittances. Absent from AWS Transcribe entirely and capped at Chirp 2 on Google.
49 AI readiness
33m
Across 12 countries and 5 language families — West, East, Southern, Central and the Horn — each with its resource tier, its dialects and the corpora that actually cover it, including the ones where the honest answer is that nothing does.
Positions
What we hold
How this compounds
Most of what a v1 model needs is already free — 11,000 hours from one philanthropic release, 33,150 more under CC0, an Apache-2.0 model covering 51 languages. Competing on hours is competing with Google and Gates for something they are giving away. So we do not.
A failed-transfer support agent for a leading African fintech, in regional languages, over WhatsApp and the phone. One workflow done better than anyone, not a generic platform.
Noisy, 8 kHz, code-switched, dense with account numbers, bank names and surnames. Nobody is releasing this data, because it is produced by deployments rather than by grant-funded collection.
Every correction is a labelled example. The generic layer under every tenant improves, while tenant data never enters the shared graph without contractual permission.
The second bank gets a model that already knows its competitors' names, its country's places and its customers' surnames. That is the compounding asset — not the hours.
Platform targets, not results
Every figure below is a stated goal, not a milestone reached. We would rather publish a plan you can hold us to than a number you cannot check — and the gap between the two is the whole reason this section is labelled.
Languages with a production model
5
Dialects mapped
40
Telephony-channel hours
1,000hrs
Verified speakers
3,500
Studio-grade TTS hours
400hrs
Annotated records
250,000
Consent grants on file
100%
Open cultural records
10,000
Models and datasets
pison-asr-afr-1
Speech recognition, high-density African language clusters, telephony-channel
pison-tts-yor-1
Yoruba text-to-speech, diacritic-native front end
pison-embed-1
Multilingual embeddings, 1024 dimensions
African pronunciation lexicon
Proper nouns, places and entities with tone marks and regional variants
Nothing here is released yet, and the status column says so rather than implying otherwise. When a model ships it carries a public training-data summary listing its sources, their licences and the consent scope they were collected under.
University partnerships, community language associations and grant-funded collection for languages with no commercial path. The public corpus is a public good and is licensed as one.
