Māori Text-to-Speech Model Spurns Big Tech's Values

Overview

Researchers at the University of Waikato built a high-fidelity text-to-speech system for a specific dialect of te reo Maori, designed so the synthetic voice and its training data remain owned by the people who speak that dialect. The project responds to concerns that big technology companies scraped Maori language data without permission and now control how that knowledge is transferred. The team hopes the approach offers a replicable blueprint for other minority language communities.

Key Takeaways

A team at the University of Waikato built a text-to-speech system for the Waikato-Maniapoto dialect of te reo Maori.
Every technical decision was shaped by the constraint that the voice and its data stay owned by the dialect's speakers.
The project responds to concerns that companies scraped Maori data without permission and control its use.
Te reo Maori is a low-resource language with relatively little digital training data.
The team hopes their approach offers a replicable blueprint for other minority language communities.

Stats & Key Facts

#Te reo Maori is spoken fluently by just 4.3 percent of the population
#About 30 percent of New Zealanders can speak more than a few words or phrases
#Initial recordings gave 4.5 hours of data
#The final dataset totaled 7 hours of audio

Māori Text-to-Speech Model Spurns Big Tech's Values

Why a sovereign Maori voice

Existing AI systems already speak Maori, but on others' terms.

›ChatGPT, Claude and Perplexity can already write te reo Maori fluently.
›That performance is built on Maori text and audio scraped without permission and processed outside New Zealand.
›Professor Te Taka Keegan calls for 'sovereign digital systems' owned by the community.

Keegan said overseas companies scraped the data with no input from Maori, and the community does not own the output. He framed language as the most important conveyor of knowledge, warning that outside technology gains more control over the transfer of that knowledge.

The team and the constraint

Ownership shaped every choice.

›Keegan and his master's student Kingsley Eng set out to develop the synthetic voice.
›The foundational constraint was that the voice and all data must remain owned by the people who speak the dialect.
›Keegan is a professor at the University of Waikato and codirector of its Artificial Intelligence Institute.

Linguistic challenges

Te reo Maori has features that trip up English-built models.

›AI voice models are predominantly built in English, which leads to errors in other languages.
›Vowel length changes meaning: keke (cake), keke (armpit) and keke (to creak) differ only by vowel length.
›Digraphs are common and pronounced differently than in English, with 'wh' usually pronounced 'f.'

Te reo Maori is considered a low-resource language because, compared with English or Chinese, there is relatively little digital training data in text, datasets or recorded speech.

Building the dataset

The team focused on a local dialect with a consenting voice.

›They chose the Waikato-Maniapoto dialect to capture the beauty tied to a specific place and identity.
›Ngaringi Katipa, a translator, educator and language mentor, was the consenting human voice.
›Initial recordings of book passages gave 4.5 hours of data.

The dataset was later expanded with a comprehensive list of sentences and words, including very rare words, supplied by Keegan's brother Peter, a Maori linguistics expert. Once cleaned and processed, the final tally was 7 hours of audio.

A blueprint for others

›The team hopes their work offers a replicable blueprint for other minority language communities.
›The approach keeps data ownership with the language community.
›It addresses the gap left by English-centric AI voice models.

Frequently Asked Questions

Who built the Maori text-to-speech system?

Te Taka Keegan, a professor at the University of Waikato, and his then master's student Kingsley Eng.

What makes this project different?

Every technical decision was shaped by the constraint that the synthetic voice and all data used to build it remain owned by the people who speak that dialect.

Why is te reo Maori hard for AI voice models?

It is a low-resource language with features like meaningful vowel length and digraphs, and most AI voice models are built in English.

How much data did the team collect?

Initial book recordings gave 4.5 hours, and the final cleaned dataset totaled 7 hours of audio.

Who provided the voice?

Ngaringi Katipa, a translator, educator and language mentor, served as the consenting human voice behind the tool.

The Waikato team hopes its community-owned approach to a Maori voice model can serve as a blueprint for other minority languages.