Te Hiku Media's Papa Reo project (NZD 13 million, MBIE 2019–2026) built AI speech recognition (92% accuracy) and a pronunciation app for te reo Māori learners, governed under an indigenous data-sovereignty (kaitiakitanga) framework cited globally as a replicable model.
non-profit, indigenous media, language technology, speech recognition, accessibility
Context
Papa Reo is a seven-year (2019-2026) data science project funded by New Zealand's Ministry of Business, Innovation and Employment (MBIE) with NZD 13 million, led by Te Hiku Media — the media organisation of Te Rarawa and Ngāti Kuri iwi in the Far North (Northland) of New Zealand. The goal is to harness AI to revitalise te reo Māori, an endangered indigenous language.
Objectives
To build AI speech technology — automatic speech recognition, text-to-speech synthesis and a bilingual transcription tool — that supports te reo Māori learners and revitalises the language, while keeping all Māori voice data under community control through an explicit data-sovereignty framework.
Activities
Papa Reo developed the first automatic speech recognition (ASR) model for te reo Māori, achieving 92% transcription accuracy using 300 hours of community-annotated audio and NVIDIA's NeMo toolkit. A text-to-speech (TTS) synthesis system and a bilingual te reo/English transcription tool (82% accuracy) were also built; the TTS work was published at ACL 2024. The flagship learning application is the Rongo pronunciation app, guiding users through 24 progressive lessons covering Māori phonemes, with real-time, on-device feedback comparing the learner's pronunciation to the ASR model, enabling practice without internet connectivity.
Results
NZ Herald and BDO NZ have documented community uptake of the Rongo app. MIT Technology Review (April 2022) and NVIDIA cited Te Hiku as an international template for indigenous-governed AI. Caution: no controlled study on te reo Māori learning outcomes has been published to date; evidence rests on ASR technical performance, government funding commitment, and community adoption of apps. Broader impact data are expected as deployment in kura (Māori-medium schools) expands.
Conclusions
A defining feature is the kaitiakitanga data-sovereignty framework: all Māori voice recordings remain under community control with explicit consent protocols, so the trained models cannot be commercialised without iwi approval. The Papa Reo architecture is explicitly designed for transfer to other under-resourced indigenous language communities.
Implementation
Indicative cost
High (€500k–€5M) — NZD 13 million over the seven-year (2019-2026) MBIE grant period.
Time to results
Long (> 3 years) — Seven-year funded project (2019-2026); ASR and TTS systems already built and published (ACL 2024), with broader school (kura) deployment described as still expanding.
Staffing & skills
Te Hiku Media (Te Rarawa and Ngāti Kuri iwi media organisation) as lead, community members recording/annotating audio, MBIE-funded project team
Conditions for success
kaitiakitanga data-sovereignty framework with explicit community consent protocols for all voice recordings
sufficient community-annotated audio (300 hours) to train the ASR model
on-device operation so learners can practise without internet connectivity
Common failure modes
accuracy and coverage are bounded by the amount of community-annotated audio available
no controlled study on learning outcomes yet published, limiting evidence beyond technical/adoption metrics
Where it fits
Governance type
iwi-led media organisation with government (MBIE) funding
Scale
regional to national (Aotearoa New Zealand); designed for transfer to other indigenous-language communities
Income level
high-income (New Zealand)
Data sources
Where this practice's information was retrieved from, and when.
STMicroelectronics Foundation, ITU and a blind Milan professor run train-the-trainer digital-literacy classes for blind and visually-impaired adults in Dakar. An …
