Whose voice is Aymara?
Building Aymara speech interfaces with communities, and who owns the recordings
This event is part of a demonstration of what Lera offers for running knowledge events. Its content is written with care and draws on real sources, but the conference has not taken place, and apart from the convener its participants are fictional.
Aymara is the language of millions of people in Bolivia, Peru and Chile, and Bolivia's 2009 constitution recognises it as an official language alongside dozens of other Indigenous languages. But official status on paper does not turn into tools: the rural clinic, the local radio station and the civil registry still work in Spanish. For the past two years this project has been building, in several communities on the shores of Lake Titicaca and in the city of El Alto, a simple speech interface for health messages and radio programmes that can understand and answer in Aymara.
The hardest part of the work was not technical. Training any speech system takes hundreds of hours of audio, and the voice of an elder telling the story of planting potatoes is not "data"; it is knowledge that has an owner. Before any recording, the assembly of each community decided several things: what would be recorded, where the recordings would be kept, who would have the right to use them, and what the answer would be if a company one day wanted to buy the model. The CARE principles for Indigenous data governance, formulated by Carroll and colleagues in 2020, framed this conversation.
The result runs against the prevailing norm of "open data". Our corpus is not open, but it is not hidden either: anyone who wants to use it asks the assembly for permission and says where the benefit of their work will return. Bird (2020) has shown that language technology for Indigenous communities often repeats the same extractive logic that endangered those languages in the first place. We tried to set that logic aside from the very first day.
Findings
- When the community holds the decision from the start, participation in recording is greater and lasts longer; no assembly left the project.
- Open data and justice are not the same thing; sometimes "closed but accountable" is fairer than "open to all".
- A small local system that people own serves the clinic better than a global model that also "supports" Aymara.
References
Carroll, Stephanie Russo, et al. 2020. "The CARE Principles for Indigenous Data Governance." Data Science Journal 19 (1): 43.
Bird, Steven. 2020. "Decolonising Speech and Language Technology." Proceedings of the 28th International Conference on Computational Linguistics (COLING 2020), 3504–3519.
Constitución Política del Estado Plurinacional de Bolivia. 2009. Article 5.