Speech-based generative AI for low resource languages: Promises and pitfalls

Abstract

Existing collections for spoken languages have numerous barriers with which speech technology can assist. In sharing forums, Indigenous Australian community members often raise orthography as a barrier in language work. For some Native American communities, text to speech systems (TTS) have proved useful. We compared TTS model outputs for one Australian language, and in doing so, ended up creating a deepfake of an elder. Existing guidelines do not adequately cover the possibilities and concerns raised by this area of generative AI and protocols for the use of archival materials for this type of speech technology are urgently needed. We suggest some preliminary guidelines.