Uyari

Uyari ("listen" in Quechua) is an open toolkit that builds translation models, in text and speech, for any Indigenous language, from the community's own human translations. It was built during the hackathon on the Meta Model API (Muse Spark, Muse Voice Transcribe) and Meta's open-weight models (Muse Glimmer 30B, NLLB-200, MMS-TTS). THE PROBLEM: the language gap in basic services. 476 million Indigenous people live in 90 countries (ILO 2019), and many face doctors, consent forms, courts and civil registries in a language they don't fully understand. In just four Latin American countries, 3.1 million people can't use the official language at all (INEGI 2020, Instituto Cervantes 2024), and 70% of Indigenous people live in Asia and the Pacific, where no one has counted. In Guatemala, language was the most important obstacle to Indigenous women's reproductive health care, and their maternal mortality was 166 vs 78 per 100,000 live births (Guttmacher). Peru has 859 registered interpreters for 4.5 million Indigenous mother-tongue speakers (Ministry of Culture 2025). Commercial translators skip these languages because there is no data. WHO BENEFITS: the person who needs a diagnosis, a consent form or a hearing in their own language, and the health workers and interpreters who serve them. HOW IT WORKS: machines write the official-language side, and human translators write the Indigenous side. 1. Muse Spark writes Spanish sentences from a design matrix of everyday life. 2. Muse Spark translates them into seven more languages, keeping each sentence's id. 3. Commissioned translators or community teams translate them into the Indigenous language. 4. Open Meta models are fine-tuned on the result and published. A new language is a folder of translations, and nobody changes code. Consent templates and the rules on rights and withdrawal come in eight languages. FIRST CASE: Shawi, with 17,241 mother-tongue speakers in the Peruvian Amazon (2017 census). The work began with maternal health. From 3,575 human translations, Muse Glimmer 30B learned Shawi ↔ Spanish, English, Portuguese, French, German and Italian. TRY IT LIVE: https://uyari.ngrok.dev - Type Spanish: the Shawi comes back in about 0.6 s, read aloud, with the three closest human translations underneath so a speaker can check it. - Type Shawi to get it in any of the six languages. - Speak, with your own Meta key. Made before the hackathon, and ours: 10,000 Spanish sentences (June 2026) and 3,578 human Shawi translations of them (July–October 2026). Everything else was built during the hackathon, 26 October – 4 November 2026, as the README discloses.