evidoria

← Back to browse

Good practice Imported

Digital Umuganda — Building Open Voice Data and AI Services for Kinyarwanda, Rwanda's National Language

Rwanda · Kigali · See the Rwanda profile · See the Kigali profile

Evidence: Descriptive / self-reported Top 24% 92/100 · Ask Evidence Copilot about this practice

Kigali-based Digital Umuganda crowdsourced one of the largest open Kinyarwanda voice datasets (1,700+ hours from 840+ contributors) via Mozilla Common Voice, then built Mbaza, a Kinyarwanda voice chatbot that delivered COVID-19 information with GIZ.

1,700+ hours
Kinyarwanda speech hours crowdsourced (by May 2021)
840+
Contributors (by May 2021)
12,000,000+
Kinyarwanda speakers (language reach)
Digital Umuganda — Building Open Voice Data and AI Services for Kinyarwanda, Rwanda's National Language

Details

Maturity
Established
Promoter
Digital Umuganda
Period
2019–present
Keywords
voice technology, open data, public health communication, language technology

Context

Digital Umuganda grew out of a 2019 Kigali hackathon with Mozilla, addressing the near-total absence of Kinyarwanda voice data (spoken by over 12 million people) on major speech-AI platforms, funded through Germany's 'FAIR Forward -- AI for All' programme (BMZ/GIZ).

Objectives

The project aimed to crowdsource an open Kinyarwanda speech dataset via Mozilla's Common Voice project and use it to build practical Kinyarwanda-language voice AI services.

Activities

Community recording drives grew the Kinyarwanda Common Voice dataset to more than 1,700 hours from over 840 contributors by May 2021; the team then built 'Mbaza,' a Kinyarwanda voice-and-text chatbot with GIZ delivering COVID-19 health information, and has continued developing speech-to-text infrastructure with releases through 2025.

Results

The Kinyarwanda dataset became one of Common Voice's largest non-English datasets, and Mozilla published a dedicated lessons-learned account covering challenges such as standardising Kinyarwanda orthography for transcription and balancing speaker samples by age, gender and dialect.

Conclusions

The project shows sustained, multi-year technical continuity (2019-2025) and rare documented-learning transparency, though most reach figures beyond the Mozilla-hosted dataset size are self-reported by partner organisations rather than independently audited.

Implementation
Indicative cost
Medium (€50k–€500k) — Donor-funded (GIZ/BMZ FAIR Forward) crowdsourcing campaign and AI development team; moderate ongoing cost given reuse of an open-source data platform.
Time to results
Medium (1–3 years) — From 2019 hackathon to a 1,700+ hour dataset by 2021 and continued technical releases through 2025 (6+ years).
Staffing & skills
Digital Umuganda team (CEO Audace Niyonkuru), community volunteer voice contributors (840+), GIZ/BMZ FAIR Forward technical and funding partner, Mozilla Common Voice platform team

Conditions for success

  • large-scale community crowdsourcing of voice recordings
  • donor technical-assistance funding (GIZ/BMZ)
  • partnership with an existing open-data platform (Mozilla Common Voice) rather than building infrastructure from scratch
  • attention to linguistic issues (orthography standardisation, speaker demographic balance)

Common failure modes

  • standardising Kinyarwanda orthography for transcription was a documented recurring difficulty
  • most downstream reach/impact figures are self-reported by partner organisations

Where it fits

Governance type
local tech non-profit/social enterprise with international donor partnership
Scale
national language (Kinyarwanda, 12M+ speakers)
Income level
low-income

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful