Consented provenance
Every single voice in our corpus was recorded with informed consent and fair, direct economic remuneration. We reject web-scraped pirated speech data. Linguistic communities own their voice heritage.
Loading
About Pison Labs
2,000+
languages & dialects natively spoken across the African continent
1.4B
people whose native speech remains underserved by global tech
100%
consented, community-recorded clean training corpora
Global frontier AI models were trained on Western audio and European text. They fail on African tonal systems, distort local names and currencies, and collapse over standard 8 kHz phone lines. Pison Labs is building the foundational speech recognition, text-to-speech, translation, and voice agent infrastructure engineered specifically for the realities of African communication.
Our thesis
The Linguistic Reality
Frontier AI models score impressive numbers on clean, studio-recorded evaluations. But when deployed in African commercial workflows — customer call centers, mobile banking, remittance confirmation, or healthcare triage — they run into failure modes no global benchmark measures.
In African languages, tone is not musical expression; it determines grammatical meaning. Stripping diacritics in Yoruba or Igbo during tokenization fundamentally alters sentences. Furthermore, African speakers continuously code-switch between indigenous languages, English, French, and regional Pidgins.
Most importantly, African business happens over telephone lines. Narrowband audio (8 kHz) contains less than half the spectral information of studio speech. An AI model that cannot understand an accented call with background traffic noise is not a product; it is an experiment.
Standard NLP pipelines strip accents and diacritics during preprocessing. In Yoruba, the word “ọwọ” can mean hand, respect, or broom depending entirely on high, mid, or low pitch accents. Pison maintains tone-aware phonemic representations across every stage.
Every contact center in Lagos, Nairobi, and Johannesburg operates on narrowband cellular lines. Pison models are trained on real-world telephony impairments, codec compression, and line jitter, achieving enterprise WER under production conditions.
When a customer speaks an account number, Nigerian bank name, or Hausa surname, standard models fail 40–70% of the time. Our entity lexicon and custom acoustic decoders guarantee precision on the words that decide financial transactions.
Foundational principles
How We Operate
Every single voice in our corpus was recorded with informed consent and fair, direct economic remuneration. We reject web-scraped pirated speech data. Linguistic communities own their voice heritage.
African commerce and contact centers operate over 8 kHz cellular lines, USSD, and WhatsApp voice notes — not studio microphones. Our models are built to excel where acoustic conditions are noisy and bandwidth is scarce.
In tonal African languages, tone is grammatical, not decorative. Stripping diacritics changes words into entirely different meanings. Our tokenizers and acoustic models preserve tonal contours end-to-end.
We publish independent benchmarks, open language metrics, and transparent evaluation criteria. True technological advancement in African NLP requires verifiable results over marketing claims.
Try all speech, translation, and agent capabilities directly in your browser with zero setup, or speak with our engineers about dedicated enterprise models.