Eswatini • Language • AI • Cultural Preservation

AI can't pronounce
siSwati correctly.

siSwati is a tonal language — the same written word can mean completely different things depending on pitch. AI systems have no way to know the difference. We're fixing that.

The pronunciation problem, live

Press play on each card to hear how a generic AI pronounces these siSwati words — then see what the correct pronunciation should convey. The same spelling, two different meanings.

* Examples below are illustrative. Exact diacritic notation is pending linguistic expert validation.

Example 1
AI reads Wrong
bana
tone changes meaning
With diacritics Clear
bána
High-High tone → "children"
banà
High-Low tone → "theirs"
Native recording coming soon
Example 2
AI reads Wrong
lifa
tone changes meaning
With diacritics Clear
lífa
Rising tone → "inheritance"
lifà
Falling tone → "to die"
Native recording coming soon
Example 3
AI reads Wrong
sita
tone changes meaning
With diacritics Clear
síta
High tone → "to help"
sità
Low tone → "enemy"
Native recording coming soon

What you just heard is the problem

The AI above pronounced siSwati using an English phonetic model. It has no concept of Bantu tonal structure. The resulting speech is not only mispronounced — in real communication, it would convey the wrong meaning entirely. This affects every siSwati AI assistant, voice tool, and educational application built today.

Why this matters at scale

Research on comparable orthographic reforms in tonal languages projects significant gains across literacy, comprehension, and digital inclusion.

Literacy Rate +15%

Projected over 5-year implementation. Based on comparable orthographic reforms (UNESCO, 2016).

Reading Speed +25%

Tone marking improved reading speed 27% in comparable Bantu languages (Bird, 1999).

Reading Comprehension +35%

Orthographic clarity → 30–40% comprehension improvement (Roberts et al., 2022).

Digital / AI Adoption +40%

Standardised orthography improved NLP performance 35–50% for African languages (Adebara, 2022).

Aligned with UN Sustainable Development Goals

SDG 4 • Quality Education SDG 9 • Innovation & Infrastructure SDG 10 • Reduced Inequalities SDG 11 • Sustainable Communities SDG 16 • Strong Institutions SDG 17 • Partnerships

What we're building

A phased initiative to standardise siSwati diacritics and integrate them into AI speech systems — starting with the words where getting it wrong matters most.

Phase 1 • 0–6 months

Research & Word Collection

Document ambiguous words, establish diacritic notation standards with linguists, begin native speaker recordings.

Phase 2 • 6–12 months

Stakeholder Engagement

Align with Ministry of Education, University of Eswatini, LiSwati Lekubhala, UNDP, and UNESCO.

Phase 3 • 12–18 months

AI Integration

Fine-tune TTS/STT models on diacritic-annotated siSwati datasets. Demonstrate measurable accuracy improvements.

Phase 4 • 18–24 months

Education Rollout

Deploy literacy tools, reading applications, and teacher resources using the new diacritic-enabled AI models.

Phase 5 • 24–30 months

Scale & Policy

Formalise diacritics in national orthography standards. Open-source the dataset and models for other Bantu languages.

Help us teach AI to speak siSwati

Every voice recording from a native siSwati speaker gets us closer to AI that pronounces words correctly, preserves meaning, and gives every Swati speaker access to the digital world on their own terms.

Native speakers

Record yourself reading siSwati words to help build our training dataset.

Linguists

Help validate diacritic notation and curate the ambiguous word list.

Partners

Schools, government, and NGOs can integrate early tools and provide feedback.

Get involved → Voice recording app launching soon