Indigenous Language ASR Is Not a Recorder: Before Training Elder Voices into Models, Let's Discuss Who Has the Right to Remain Silent

# Indigenous Language ASR Is Not a Recorder: Before Training Elder Voices into Models, Let's Discuss Who Has the Right to Remain Silent

Putting voices into models is not cultural revitalization

Many people hear about Indigenous language ASR for the first time and immediately show engineer-style excitement: great, put elder recordings into a model and it will automatically transcribe them; then add translation, then speech synthesis, then a chatbot, and cultural revitalization is complete. This imagination is efficient but also dangerous because it treats language as sound, sound as data, data as fuel, and finally Indigenous people as fuel suppliers.

Indigenous language is not an audio file. It is relationship.

The meaning of a word may be hidden in who says it to whom, when (season), where (place name) and whether it can be spoken publicly. An elder's narrative may simultaneously contain family memory, migration experience, place names, taboos, jokes, bodily sensations, historical trauma and teaching styles. If an ASR system only asks "what is the word error rate?" it will be like measuring the height of ancestors with a caliper: precise but absurd.

Low-resource does not mean low ethics

Low-resource language technology certainly needs data. Without enough speech, transcriptions, corpora and annotations, models struggle to recognize phonemes, stress, intonation and context. The problem is that insufficient data does not mean we can skip consent; Indigenous languages are precious but not all of them should be opened up.

The CARE principles remind us that data governance cannot only talk about findable, accessible, interoperable and reusable (FAIR); it must also address collective benefit, control, responsibility and ethics. This is a very basic reminder: data does not float in the air; it comes from people and will affect them.

Taiwan Indigenous language AI faces the same issues. Today we want to build Indigenous language ASR, speech dictionaries, chatbots, textbook generation tools—directions worth affirming. But truly mature systems are not about who collects the most audio files; they are about designing clear governance rules: which data can be public? Which limited to community internal use? Which only for teaching and not commercial? Which involve rituals, families or places that need to be sealed? Who can authorize? Who can withdraw? Who can query usage records? How do benefits return to the community?

Slowness is the ethical speed Indigenous language technology must keep

These questions are slow. Slow enough that they don't fit on a tech conference stage because they cannot become three-minute demos. But what Indigenous revitalization fears most is using demo speed to process memory. The black humor is: tech companies can spend two years making models smarter, yet complain that spending two months opening community meetings is inefficient.

Designing Indigenous language ASR should have at least four layers. First the technical layer: recording quality, transcription standards, phonological analysis, model evaluation. Second the educational layer: enabling teachers and learners to actually use it, not just for research reports. Third the governance layer: community authorization, data classification, withdrawal mechanisms, usage records and benefit feedback. Fourth the cultural layer: acknowledging that some voices should not be learned by models; some silences are themselves knowledge.

The right to remain silent is also data sovereignty

"The right to remain silent" is the most overlooked right in Indigenous language AI. Modern data civilization likes to treat silence as missing data, and AI loves to fill blanks. But in many cultural contexts, silence is not a lack of information; it is a boundary. When an elder does not speak, it may not be ignorance; it may mean you are not yet qualified to listen, or that the content should not be spoken publicly. A truly mature AI system should pause at such places rather than enthusiastically generating a culturally flavored nonsense.

Indigenous language technology can still be done; precisely because it is important, it must not be done roughly. Good ASR can assist teaching, preserve pronunciation, support textbook production and enable dispersed Indigenous people to reconnect with their mother tongue. But it must stand under community governance rather than turning the community into a model's material library.

Before machines understand, let the community decide

If future Indigenous language AI succeeds, it will not be because machines understood Indigenous languages; it will be because Indigenous peoples still have the right to decide how their voices are heard. True revitalization is not letting machines speak for us; it is ensuring that the next generation, alongside technology, still knows when to open their mouths and when to remain silent.

We need technology, but also an ethics that can make technology pause. The best future of Indigenous language ASR is not turning all voices into data; it is allowing each voice to be heard, preserved or properly left unsaved within the right relationships.

Further reading and sources

• Global Indigenous Data Alliance: CARE Principles
• Indigenous Data Sovereignty literature
• Māori data sovereignty and language technology cases