Indigenous Language AI Is Not a Keyboard Issue: When Voice Data Becomes Public Service Infrastructure
Original Chinese title: 族語 AI 不是鍵盤問題：當語音資料變成公共服務基礎建設
The core of Indigenous language AI is not building another input method, but determining who has the authority to decide how voice data is collected, verified, preserved, licensed, and returned to public service contexts.
Two-Eyed Seeing Lab
An editorial team focused on Indigenous Peoples' knowledge, cultural data sovereignty, and AI collaborative design.
Before asking whether a model supports it, ask who has the authority to decide
Every time Indigenous language AI is discussed, the tech community tends to reduce the issue to three things: keyboard support, speech recognition capability, and translation ability of the model. These are certainly important, but if we stop there, Indigenous languages risk being downgraded to mere engineering functions—as though language were just a set of symbols that can be placed in datasets, run through training pipelines, and connected via APIs, thereby completing revitalization. This mindset is convenient yet dangerous.
Indigenous languages are not simply "multi-language support" for general products. They involve family memory, land relationships, ritual taboos, intergenerational trauma, educational systems, and accessibility to public services. When a recording of an Indigenous language is used to train a model, it is more than just audio files; it may embody the knowledge of an elder, the oral context of a community, or vocabulary appropriate only for specific occasions. If engineers see only file length and transcription accuracy, that is not progress—it reflects too narrow a perspective.
Voice data is not a found seashell
Open voice databases are highly beneficial for low-resource languages. Without sufficient audio, automatic speech recognition cannot function; without adequate transcriptions, voice assistants will never understand mainstream languages beyond the dominant ones; without testable datasets, government services struggle to become truly multilingual and equitable. However, "open" does not automatically equal justice. For Indigenous Peoples' languages, openness must answer: who consents? who governs? who benefits? who can withdraw consent? which content should never be opened?
The CARE principles remind us that data governance must go beyond FAIR to recognize collective interests, control, responsibility, and ethics. In other words, Indigenous language data cannot be designed solely for ease of model use; it must also be structured so communities retain authority over their own knowledge. If an Indigenous language AI ends up allowing external companies to take the data, governments to claim the results, and academic institutions to publish papers—while the community receives only a showcase presentation—that is not digital revitalization; it is cloud-based data colonialism.
Public services need Indigenous languages, but they must not steal them
The truly worthwhile Indigenous language AI projects are not about showing off technical prowess, but serving public needs. Can elderly patients describe pain in their familiar language when seeking medical care? Can disaster alerts convey evacuation and shelter instructions clearly in Indigenous languages? Can schools enable children to hear their own language as a knowledge tool for research, questioning, and homework—not merely as performance? These questions are not romantic; they are practical.
Yet public services are also the most vulnerable to efficiency overriding consent. Governments may argue that centralizing data is necessary to serve more people; vendors might claim that higher accuracy requires additional recordings; researchers could contend that preserving language demands collecting everything possible first. All these reasons sound benevolent—so benevolent that we must be especially vigilant. Once cultural data enters a system, it will be copied, backed up, converted, and repurposed. What is promised today as speech recognition may become voice generation tomorrow, or commercial customer service the day after. Indigenous languages are not ownerless mines.
Two-Eyed Seeing is not merely connecting communities to platforms
The essence of Two-Eyed Seeing lies not in "technologizing" community knowledge, but in re-educating technology itself through local ethics. Designing Indigenous language AI must begin with governance: establishing shared norms before data collection; ensuring recorders understand purpose and risks; enabling tribal or ethnic organizations to participate in licensing; creating exclusion mechanisms for sensitive vocabulary and taboo knowledge; providing channels for reporting model errors; and returning outcomes to education, healthcare, disaster preparedness, and cultural transmission contexts.
This approach will be slower. Slowness is not necessarily a drawback. Language revitalization has never been a hackathon weekend demo that earns awards on stage. It resembles tending a sacred fire: knowing its origin, who may approach it, when to add fuel, and when to let it rest. AI can assist, but it must not turn the fire into slide-deck material.
Ultimately, Indigenous language AI must ask not "Will the model speak?" but "When the model begins speaking, do community members retain decision-making authority?" If the answer is no, even the most polished voice interface becomes another form of precise accent-based plunder. If yes, Indigenous language technology can evolve from an engineering project into genuine public infrastructure—standing firmly on the side of cultural subjectivity.
Technology can accelerate, but it cannot replace relationships
The greatest challenge in Indigenous language public services is not translating a button; it is ensuring users trust the system when they truly need help. Elderly patients describing pain in medical settings require more than high recognition rates—they demand awareness that errors could compromise care. Disaster alerts broadcast in Indigenous languages must convey clear evacuation locations, times, transportation routes, and responsible agencies—not just catchy slogans. Once language technology enters public services, it engages lives, dignity, and rights.
Therefore, Indigenous language AI requires not a single product launch but an auditable system: where data originates, who may use it, how errors are reported, how re-licensing occurs during model updates, and how outcomes return to schools and communities. Technology can accelerate revitalization, yet revitalization itself remains relational engineering. Without relationships, only models remain; without governance, only showcases persist.
Sources retained from the Chinese original
AI use and content-safety disclosure
This article was assisted by AI for data organization, structural drafting, and sentence polishing; human editors set the viewpoint and fact-checking direction