Māori communities build their own AI voice models — on their own terms

Much of the debate about AI and Indigenous languages has been shadowed by a fear of “data colonialism” — outside companies scraping a community’s words to build tools the community neither controls nor benefits from. A report published on May 21, 2026 describes a hopeful alternative taking shape in Aotearoa New Zealand, where Māori researchers and organizations are building AI that speaks te reo Māori on their own terms.

One project comes from the University of Waikato, where researcher Te Taka Keegan and colleague Kingsley Eng built a text-to-speech tool for the Waikato-Maniapoto dialect. The synthetic voice is based on recordings by Ngaringi Katipa, a translator and language mentor who consented to be its human voice. Trained on just under eight hours of audio and built with a phoneme-based approach and the Piper neural architecture, the system reached a 6.78% word error rate — considered good by industry standards — and in listening tests native speakers could tell synthetic from human audio only about 65% of the time. Crucially, guardianship of the tool sits with three iwi (tribes): Waikato, Maniapoto and Raukawa.

“A report published on May 21, 2026 describes a hopeful alternative taking shape in Aotearoa New Zealand, where Māori researchers and organizations are building AI that speaks te reo Māori on their own terms.”

A second effort, Te Hiku Media’s Papa Reo, tackles the reverse problem: automatic speech recognition, transcribing spoken te reo Māori with about 92% accuracy, and around 82% for bilingual speech. Te Hiku Media released its work under a “Kaitiakitanga” license — a term rooted in the Māori concept of guardianship — that restricts use of the models and data so they benefit Māori people rather than being exploited by outsiders.

“Our language is the most important conveyor we have for our knowledge, yet we see technology developed outside of Aotearoa get more and more control over the transfer of that knowledge,” said Te Taka Keegan — a line that captures why community ownership, not just accuracy, is the point.

What makes this a genuine good-news story is not that the technology is flawless — these are early tools trained on modest data, and revitalizing a language takes far more than software. It is the model of how the work is being done: efficient training on small datasets, phoneme-based inputs, open-source tools and, above all, a governance framework in which the community owns both the data and the models. The researchers have deliberately laid a template that other iwi — and Indigenous peoples elsewhere — can follow, turning AI from a threat to endangered languages into a tool their own speakers command.