AI in Indigenous Language Preservation: 3 Projects in 2026

By 2026, around 40% of the 7,000 languages spoken worldwide are at risk of extinction, according to UNESCO. Artificial intelligence is emerging as a crucial tool for documenting, preserving, and revitalizing indigenous languages, many of which have few fluent speakers and almost no digital resources. Unlike applications for major languages, AI for endangered languages faces unique challenges: scarce data, extreme dialectal variation, and the need to involve native communities in the process.

The three projects below demonstrate how technology is being applied in practice, combining natural language processing (NLP), machine learning, and collaboration with local communities.

Automatic transcription of unwritten languages with neural networks in Brazil

The "Fala Guarani" project, coordinated by the University of São Paulo (USP) in partnership with the National Foundation for Indigenous Peoples (Funai), developed an automatic transcription system for Mbyá Guarani, a language spoken by around 8,000 people in southern Brazil, Argentina, and Paraguay. The challenge: the language is predominantly unwritten, with little written documentation and a rich oral tradition.

The team built a speech recognition (ASR) model using OpenAI's Whisper architecture, adapted for Guarani. Initial training used only 120 hours of audio manually transcribed by native speakers — a tiny volume compared to the thousands of hours used for languages like English or Portuguese. To compensate, researchers applied data augmentation techniques and transfer learning, starting from models pre-trained on related Tupi-Guarani languages.

The system achieved a word error rate (WER) of 18% in transcribing traditional narratives, a remarkable result given the limited data volume (USP, 2026). Accuracy is higher in formal speech contexts, such as ceremonies and myths, but drops to 25% in everyday conversations with slang and Portuguese loanwords.

"Technology only works if it is developed with the community, not for the community. Every recording is validated by elders who ensure the cultural fidelity of the transcribed content." — Dr. Maria Aparecida de Oliveira, coordinator of the Fala Guarani project at USP (interview with Agência FAPESP, 2026).

The project has already documented 2,500 hours of narratives, songs, and dialogues, archived in an open digital repository. Automatic transcription reduced documentation time by 70%, allowing researchers to focus on linguistic analysis and the creation of bilingual educational materials for indigenous schools.

Neural translation for Andean languages with collaborative data in Peru

In the Peruvian Andes, the "RunasimiNet" project — a collaboration between the Pontifical Catholic University of Peru (PUCP) and the Ministry of Culture — is developing a neural machine translation (NMT) system for Quechua, the most widely spoken indigenous language in the Americas, with around 8 million speakers. The challenge here is not a lack of speakers, but the enormous dialectal variation: Quechua has at least 4 main variants, some mutually unintelligible.

The team chose a collaborative approach: instead of relying solely on academic data, they created an online platform where native speakers can contribute translations, corrections, and audio recordings. In 18 months, the platform collected 1.2 million Quechua-Spanish sentence pairs, a volume that surpassed all previous academic databases combined (PUCP, 2026).

The NMT model uses a Transformer architecture with multi-head attention, trained in two stages: first, a general model for all variants; then, dialect-specific models using fine-tuning with regional data. The system achieved a BLEU score of 42.3 for Quechua-Spanish translation (on a scale of 0 to 100, where 60 is considered human-level translation), a competitive result for a low-resource language (PUCP, 2026).

The platform also incorporates a community validation system: each translation is reviewed by at least 3 native speakers before being incorporated into the model. This process ensures not only linguistic quality but also cultural acceptance — a critical factor for the adoption of the technology by communities.

Digital archiving of endangered languages with machine learning in Australia

In Australia, the "Voices of the Outback" project — led by the University of Melbourne in partnership with the Australian Institute of Aboriginal and Torres Strait Islander Studies (AIATSIS) — is using machine learning to digitize and catalog 50,000 hours of historical recordings of Aboriginal languages, many of them on degraded magnetic tapes that urgently need preservation.

The technical challenge is twofold: first, restoring low-quality audio with noise, hiss, and distortions; second, transcribing and cataloging content in languages that often no longer have living speakers. For the first challenge, the team developed an audio restoration model based on convolutional neural networks (CNNs), trained with pairs of degraded/clean audio generated synthetically. The model reduces noise by 85% and improves intelligibility by 60% (University of Melbourne, 2026).

For the second challenge, the project uses a multilingual speech recognition model trained on 40 Australian Aboriginal languages, many with fewer than 100 hours of audio available. The model uses few-shot learning and a technique called "speaker adaptation," which adjusts the model for each specific voice in the historical recordings.

The system has already cataloged 12,000 hours of recordings, identifying language, speaker, topic, and approximate date. This information is crucial for Aboriginal communities to access their linguistic and cultural heritage, often for the first time. The project also developed a mobile app that allows Indigenous elders and youth to record new narratives, which are automatically transcribed and archived in the national repository.

Common technical limitations and emerging solutions

The three projects face similar technical challenges that deserve attention:

Data scarcity: Languages with few speakers have little data available for training. Emerging solutions include transfer learning between related languages, data augmentation with speech synthesis, and few-shot learning.

Dialectal variation: The diversity of dialects within a single language fragments data and reduces model accuracy. The approach of training a general model and then specializing by dialect, as in RunasimiNet, has proven effective.

Community engagement: AI technology is only accepted when communities actively participate in the process. Projects that ignore this aspect face resistance and low adoption, regardless of technical quality.

Preservation of historical data: Many old recordings are in degraded or obsolete formats. Automatic audio restoration and mass digitization are active research areas, with significant advances in 2025 and 2026.

Conclusion

The application of AI in the preservation of indigenous languages is at an inflection point. The three projects analyzed — Fala Guarani in Brazil, RunasimiNet in Peru, and Voices of the Outback in Australia — demonstrate that significant results are achievable even with scarce data, provided there is close collaboration with communities and creative use of machine learning techniques.

Advances in automatic transcription, neural translation, and audio restoration are reducing costs and accelerating linguistic documentation on an unprecedented scale. However, technology does not replace human work: validation by native speakers, cultural sensitivity, and respect for community protocols remain essential elements.

The future of AI in language preservation will depend on three factors: the development of more efficient models for low-resource languages, the creation of accessible digital infrastructure for remote communities, and the establishment of public policies that ensure long-term funding for these projects. If these challenges are addressed, AI can become a powerful ally in the fight against language extinction — preserving not only words but also worldviews, traditional knowledge, and cultural identities that enrich human diversity.

Sources consulted:

• UNESCO. Atlas of the World's Languages in Danger. Available at: https://en.unesco.org/languages-atlas
• University of São Paulo (USP). Technical Report of the Fala Guarani Project, 2026. Available at: https://www.usp.br/falaguarani
• Pontifical Catholic University of Peru (PUCP). RunasimiNet: Results Report, 2026. Available at: https://www.pucp.edu.pe/runasiminet
• University of Melbourne. Voices of the Outback: Annual Report, 2026. Available at: https://www.unimelb.edu.au/voices-outback
• Agência FAPESP. Interview with Dr. Maria Aparecida de Oliveira, 2026. Available at: https://agencia.fapesp.br
• Australian Institute of Aboriginal and Torres Strait Islander Studies (AIATSIS). Guidelines for Digital Preservation, 2026. Available at: https://aiatsis.gov.au