Saving Endangered Languages
Of about 7,800 languages in Glottolog, over 4,000 are threatened, shifting, moribund or nearly extinct, and a quarter have only a wordlist or less. A language is lost when children stop learning it, and a recording alone does not keep it alive. Missing are descriptions of poorly documented languages, children learning them at home and in school, community control of data, and tools that work from a few hours of data.
Open in the interactive tree →A language is lost when children stop learning it, usually because of schooling, migration, stigma and the pull of a few large languages. Recording and describing a language takes years of work with its speakers, and a recording does not keep the language alive. The UN declared 2022-2032 the International Decade of Indigenous Languages, with a mid-term review in 2027.
As of now
Glottolog 5.3 lists 7,839 languages, of which 984 are extinct, 1,565 threatened, 1,799 shifting, 421 moribund and 300 nearly extinct; 1,911 (24%) have a wordlist or less. Ethnologue's 28th edition (2025) counts over 7,000 living languages. AI helps at the recording end: Te Hiku Media's Māori speech recogniser reaches 92% accuracy (company-reported) and Meta's Omnilingual ASR (November 2025) covers 1,600+ languages, though only 36% of the low-resource ones reach under 10% character error. Communities set rules for their data, as in Te Hiku's Kaitiakitanga Licence.
What is missing
- Recordings and descriptions for the languages that have only a wordlist
- Children learning the language at home and in school
- Community control of data and agreed licences
- Speech and text tools that work from a few hours of data
- Shared, reliable counts of speakers and vitality
Becomes possible once solved
- Children growing up in their heritage language
- Dictionaries, grammars and archives that outlive the last speaker
- Evidence about language drawn from the full range of human languages
Open steps
- Record speakers while they live Medium AI leverageDocument moribund and nearly extinct languages (421 and 300 in Glottolog 5.3) with audio, texts and wordlists before the last speakers die.
- Languages with only a wordlist Medium AI leverage1,911 languages (24% in Glottolog 5.3) have a wordlist or less; grammars and text collections for them need linguists and speakers.
- Children learning the language Low AI leverageTransmission through homes, schools and immersion decides survival; technology cannot replace speakers.
- Tools that work from few hours Medium AI leverageKeyboards, fonts, spell-checkers, speech recognition and translation for languages with little data; Omnilingual ASR adds languages from a few paired examples.
- Agree what 'endangered' means Medium AI leverageCounts differ by method: Glottolog 5.3 lists 4,085 languages as threatened, shifting, moribund or nearly extinct; other catalogues use other scales.
Where today's AI could help
Medium AI leverage. AI speeds transcription and builds tools for small languages, but speakers, time and community consent decide whether a language survives.
- Speed up transcription and glossing of recordings
- Align audio, text and translations automatically
- Build keyboards, spell-checkers and speech tools from little data
- Estimate speaker numbers and vitality from census and web data
Shown so far
- Te Hiku Media's Māori speech models reach 92% accuracy and have sped up transcription of an archive of about 1,000 hours of native-speaker content (NVIDIA blog, company-reported). source
- Meta's Omnilingual ASR lets users add a language from a few paired audio and text samples, with no retraining; low-resource languages with under ten hours of audio are the weak spot. source
Prerequisites
- Spoken Languageat least ~135,000 years ago?
- Sound Recording1877
- Unicode1991