Czy AI może transkrybować i tłumaczyć zagrożone języki z 6 godzinami danych ?

WARDEN wykorzystuje dwuetapowy system — najpierw transkrybuje dźwięk języka Wardaman na poziomie fonemicznym, a następnie tłumaczy na angielski — dysponując jedynie 6 godzinami danych treningowych. Przewyższa większe modele, wykorzystując inicjalizację podobnego języka oraz skompilowany słownik do tłumaczenia.

ŹRÓDŁO: arXiv:2605.13846 — Ziheng Zhang i in., 2026 — „WARDEN: Transkrypcja i tłumaczenie zagrożonych języków rdzennych przy użyciu 6 godzin danych treningowych”

Background

Recent work shows that, given around six hours of transcribed speech in an endangered language, modern speech-processing systems can produce usable transcriptions and even translations—provided those six hours are carefully selected and paired with related high-resource languages. Models that combine self-supervised pre-training on raw audio with fine-tuning on the small target set now reach word-error rates below 25% on some oral languages, and pivoting through a bridge language can yield BLEU scores of roughly 10–20 for short sentences. Zero-shot cross-lingual transfer from multilingual encoders such as w2v-BERT 2.0 or Whisper-large-v3 can cover phoneme inventories unseen in the six-hour sample, but intelligibility drops sharply for languages with fewer than ten speakers or highly tonal systems. Translation quality still lags behind high-resource benchmarks because grammatical patterns and idioms are under-represented in the small corpus, yet minimal post-editing is often enough to create basic bilingual lexicons or archival descriptions. Ongoing initiatives like the Lacuna Fund and UNESCO’s AI for endangered languages challenge are distributing small labeled corpora and pushing community-led data collection to make such approaches sustainable. Community partnerships remain essential: models trained only on outsider-collected data can encode cultural biases or mispronunciations unless validated by native speakers. At present, six hours is a rough lower bound; below that, data augmentation via synthetic voice conversion or back-translation becomes unreliable. Where ethical approval and speaker consent are secured, these techniques are already being deployed for language documentation, though they do not yet guarantee long-term revitalization.