Don't Throw Tribal Sounds into Databases: Soundscapes, AI Training and Cultural Data Sovereignty
Original Chinese title: 不要把部落聲音丟進資料庫：聲景、AI 訓練與文化資料主權
Sounds are not harmless background material; forest soundscapes, ritual rhythms, seasonal signals and family memories may involve cultural norms, data sovereignty and AI training boundaries.
Two-Eyed Seeing Lab
Co-authors: 林秀妹
The Two-Eyed Seeing Lab focuses on Indigenous Peoples' knowledge, cultural data sovereignty and public technology governance; co-author 林秀妹 is a full-time teacher at Beiyeh Elementary School, winner of the Indigenous Language Teacher Award and a vocalist, long committed to Indigenous language education, sound culture and local heritage.
Don't Throw Tribal Sounds into Databases: Soundscapes, AI Training and Cultural Data Sovereignty
Many people think of cultural data first as photos, artifacts, Indigenous language texts, ancient maps or oral history videos. Sounds often come last, as if they are just background music under images or atmospheric material in museum exhibitions. But for many Indigenous Peoples' communities, sounds have never been mere decoration. Insects singing through seasons, rivers changing timbre with rainfall, breathing rhythms in work songs, pauses and turns in elders' speech, even the pitch children learn to sing in certain spaces—all are part of knowledge. That is not simply downloadable content that can be re-edited, tagged and uploaded; it is living evidence of land relationships, life order and cultural memory.
The most misleading aspect of sounds is how they seem less sensitive. Photos show faces, artifacts involve collection and ownership, texts reveal meaning—sounds are often assumed to be harmless material. In reality, soundscapes may expose more clues. A bird call with water noise can pinpoint a location; a song links ritual and family; a tone appears only in certain social relationships. Once stripped of context, the first thing lost is not quality but relationship. Throwing tribal sounds directly into databases often does not preserve them; it severs their original cultural pathways and treats cross-sections as knowledge specimens.
AI's Appetite for Sounds Is Far Larger Than a Recorder's
Recorders used to simply keep sounds. Today, AI does more than preserve: it breaks sounds down into features, models and reusable capabilities. A recording is no longer just an archive from one collection; it may become part of future classification, generation, imitation, search, recommendation or commercial product training. This shift raises the risk level for soundscape data governance entirely. Once audio files train models, the issue is not only 'who can listen' but 'who can learn something through them'.
This is where many Indigenous Peoples' communities must be most alert to AI. A group's singing style may be a tiny voice dataset for platforms; a forest recording might be a biodiversity sample for external researchers; an elder's narrative could be a voice style material for generative systems. But for the community itself, these are not just sounds—they mark rights, responsibilities and taboo boundaries. AI loves to claim it only learns patterns, yet in cultural worlds many patterns should not be taken by unfamiliar systems without condition. Worse, after learning them, AI may generate music, meditation soundscapes, game backgrounds or voice interfaces that 'sound Indigenous', leaving the original users to pay for their own echoes. This is not a technological miracle; it resembles a beautifully packaged cultural boomerang.
Data Sovereignty Is Not Refusing Preservation, It Is Refusing Others' Lazy Preservation
When discussing cultural data sovereignty, many misunderstand it as 'nothing can be touched'. The real issue is not preservation itself but who decides its purpose, methods, permissions and consequences. External projects often say 'we are helping you preserve', politely and gently, yet may follow a one-way workflow: researchers decide what to record, institutions where to store, platforms who can use, papers how to interpret—communities appear once in acknowledgments. Such preservation is not without goodwill, but if that goodwill does not yield control, it remains a civilized, patient yet still colonial practice.
The CARE principles offer an important reminder here. For Indigenous Peoples' data governance, we cannot talk only about FAIR (findable, accessible, interoperable, reusable); we must also discuss Collective Benefit, Authority to Control, Responsibility and Ethics. Data is not automatically legitimate because it can flow; it must return collective benefit, retain community control, have someone responsible for outcomes and meet ethical boundaries. Soundscapes are especially so: often mistaken as 'not that sensitive', they are most easily over-shared under the guise of benevolent openness.
Local Contexts and TK Labels: The Digital World Must See Cultural Rules
Recent developments in Local Contexts and Traditional Knowledge Labels provide an inspiring approach: digital objects carry community rules. These labels do not just tell 'where it comes from'; they further remind: are there seasonal limits? Is specific consent needed? Does it involve family, gender, community, sacredness or non-commercial restrictions? Simply put, platforms and users can no longer pretend all digital data have no owners, no context, no taboos.
This matters especially for soundscape governance. Not every recording suits full openness; some may display, some educate, some apply for research, some stay internal to the community, others must be marked restricted or not recorded at all. Mature soundscape projects cannot design only how to record; they must also design what not to record, what not to publish, what can be withdrawn and what never enters model training. This work looks less lively than collecting archives but is more important: boundaryless preservation quickly becomes a pretty version of plunder.
Soundscape Data Needs Layered Governance, Not One-Click Uploads
To make soundscape data truly useful public resources, the priority is not hard drive capacity but layered governance. Public display layer, educational licensing layer, research application layer, community internal layer, taboo and sensitive layer—each needs different metadata and use conditions. Recording time, place, equipment and sound source type matter, yet are insufficient; we must also note whether AI training is allowed, commercial use permitted, cross-platform transfer okay, withdrawal rights exist, or community representatives jointly review.
For a public knowledge platform like Yuan Media AI, the truly forward-looking approach is not to treat Indigenous Peoples' soundscapes as novelty content but to demonstrate responsible discussion. We can introduce tools without stealing material; talk about data governance without turning taboos into case studies; encourage local preservation while reminding that 'not recording' and 'not publishing' are also part of preservation. The core of cultural data sovereignty has never been putting everything in the cloud, but letting knowledge pass within appropriate relationships.
Two-Eyed Seeing Is Not Translation, It Is Mutual Institutional Change
True Two-Eyed Seeing does not mean translating tribal sounds into academic formats or reversing academic terms back to local languages—that is not enough. Two-Eyed Seeing means both systems must be willing to adjust themselves. Ecoacoustics can help communities track species and environmental change; local knowledge reminds researchers which sounds should not be chopped, which silences are information, which seasons forbid recording, which occasions require stricter consent. If only the community keeps adapting to external systems while those systems refuse to change methods for the community, that is not two-way—it is merely polite one-way delivery.
Ultimately, the value of soundscape data lies not in how complete it can be collected but whether it retains original relationships in the digital age. Which sounds teach children? Which serve researchers? Which stay only in certain seasons and spaces? Which must never train generative models? These questions have no universal answers, yet they are exactly where cultural data sovereignty lives on site. Returning sounds to relationships rather than just into clouds is the true starting point of soundscape governance.
Sources retained from the Chinese original
AI use and content-safety disclosure
This article was assisted by AI for data organization, structure drafting and sentence polishing; human editors set viewpoints and fact-checking directions