01
Consented local-language datasets
Ethical text and audio collections drawn from local news, literature, radio, podcasts, and everyday speech—with regional dialects represented deliberately.
All Things projectsRequest Systems Brief
L01 · Research in formation
Zimbabwe language intelligenceInclusive language intelligence
Ilwimi.
AI that speaks our reality.
Language and speech systems built around how Zimbabweans actually speak, write, work, and reason.
Named for the Ndebele word for tongue or language
The research question
Can useful AI understand Zimbabwean languages, code-switching, local knowledge, and everyday context?
Ilwimi investigates whether models can follow the natural movement between Shona, Ndebele, Zimbabwean English, and local street language—while understanding idioms, proverbs, cultural references, and unspoken social cues.
Why it matters
Language should not become a barrier to intelligence.
Mainstream AI is largely trained on Global North data. Applied here, it can be linguistically limited and culturally tone-deaf.
When systems cannot understand local accents, dialects, or mixed-language communication, people lose access to information and local builders lose the foundation for useful products in education, health, agriculture, and public services.
Ilwimi is not simply translating words. It is research into digital inclusion and the preservation of linguistic knowledge in the age of AI.
Initial research scope
Five connected tracks.
From responsible data collection to models that can work on an ordinary phone, each track addresses a condition required for locally useful language intelligence.
02
Code-switching evaluation
Benchmarks for the fluid movement between Shona, Ndebele, Zimbabwean English, and street language, including idioms, references, and context-specific reasoning.
03
Speech, translation, and voice
ASR and text-to-speech systems tuned to Zimbabwean accents and phonetics, making natural voice a practical interface for essential digital services.
04
Efficient, low-connectivity models
Quantisation, edge computing, and lightweight architectures designed for low-end phones, intermittent connectivity, and constrained power environments.
05
Cultural context and local knowledge
Local knowledge structures that help models account for social cues, history, agricultural practice, and the reasoning behind what people mean—not only what they say.
Intended outcome
A trusted language-intelligence layer for Zimbabwe.
The outcome is infrastructure that local developers, startups, and public services can build on: from voice-based farming advice and health information in Shona to customer-service systems that understand Zimbabwean English.
Open-source models
Developer APIs
Consented datasets
Evaluation benchmarks
Voice-first interfaces
On-device intelligence
Research with us
The future of AI should include the people it is meant to serve.
Expertise
Computational linguists, NLP engineers, data scientists, anthropologists, and language practitioners.
Data
Partners with consented text, audio, or linguistic corpora in Zimbabwean languages and mixed-language settings.
Operating environments
Organisations where language systems can be evaluated responsibly under real conditions and with real users.
Research with us