Translation
Bidirectional translation across 29 African languages and English or French, preserving cultural idioms, grammatical nuances, and tonal diacritics.
Loading
Language, voice, translation and cultural intelligence infrastructure for the world’s most linguistically diverse continent
The platform
Orchestration is commoditised and priced to the floor. What is not commoditised is the layer beneath it: a model that hears Mthethwa, Otieno and Oyèlaran correctly, that knows naira from shillings from cedis, and that holds all three on a phone line, mid-code-switch.
Bidirectional translation across 29 African languages and English or French, preserving cultural idioms, grammatical nuances, and tonal diacritics.
Accurate acoustic transcription engineered for 8 kHz telephony, real-world background noise, and natural multi-language code-switching.
Natural neural voice synthesis with authentic African accents, diacritic-aware phonetics, and dynamic speed, pitch, and emotion controls.
Explore a catalog of authentic African voices filtered by language, gender, and age, or create instant custom voice clones from audio samples.
Real-time identification of spoken or written African languages, regional dialects, and code-switching segments with confidence scoring.
Contextual guidance on social etiquette, hierarchical register, and respectful honorifics and address forms across African cultures.
Automated video and audio dubbing into African languages with multi-speaker voice matching, emotion transfer, subtitles, and timeline alignment.
Conversational voice and chat agents for telephony and WhatsApp, pre-configured for African banking and mobile-money workflows, local entities, and human handoff.
Enterprise
Failed transfers, unauthorised debits and mobile-money disputes are repetitive, script-resolvable and enormous in volume. They are also the conversations most likely to happen in regional lingua francas rather than standard English.
Africa processes over 50bn mobile-money and instant-payment transactions yearly. Even a 0.1% exception rate generates tens of millions of multilingual support contacts every year from failed transactions alone.
Developers
One SDK, eight endpoints, and the same models the playground runs on. Tone and diacritics are inputs to the API rather than something a text front-end strips before the model sees them.
import { Pison } from "@pison/sdk"; const pison = new Pison({ apiKey: process.env.PISON_API_KEY }); const response = await pison.translate({ text: "Welcome home", source: "en", target: "yo",}); console.log(response.output); // "Kú àbọ̀ sílé"console.log(response.confidence); // 0.94POST /v1/translate · priced per 1M characters · the same call the playground above makes.
Questions
Everything you need to know about our platform, models, and data.
For clean-audio accuracy on many languages, you should — it is free and it is good, and pretending otherwise would not survive a technical review. What it does not give you is telephony-channel performance on 8 kHz narrowband, latency under a support-call budget, named-entity accuracy on African names and regional financial institutions, an SLA, or indemnity. That is what this is.
No, and the reason is measurable. Frontier systems run 25–35% WER on African-accented English generally but 40–70% on African named entities and 35–50% on finance vocabulary. A wrapper inherits those numbers. The work is in the layer underneath: in-domain telephony audio, a pronunciation lexicon of African proper nouns and regional entities, and diacritic-complete ground truth.
Every dataset version computes its licence from the most restrictive licence among its sources, rather than asserting one. That is what mechanically prevents publishing a CC-BY-SA and CC-BY-NC mixture, which is unpublishable under either. Non-commercial and ShareAlike corpora are used for evaluation and never pooled into a commercial training manifest.
Withdrawal is honoured within 30 days and removes the speaker from every future dataset version. We will also tell you the part most vendors do not: a model already trained cannot unlearn a voice. The consent chain runs grant to speaker to utterance to dataset version to model build precisely so that question has an answer at all.
Not yet. This design system ships one palette — a warm near-white page with a near-black brand — and a second one is a real piece of work rather than a toggle: every token, every chart step and every inverted band has to be re-derived and re-checked for contrast. Doing it badly would be worse than not doing it.
The ones people turn up to record in. Contributor sign-ups and dialect self-identification are a better demand signal than any desk estimate — one community effort put 1,211 hours of Kinyarwanda into an open corpus from 420 contributors, which no top-down prioritisation would have predicted. The catalogue ranks itself.
Start in the playground, take an API key when you are ready, and talk to us when a deployment needs private hosting or an SLA.