Leveraging Speech Recognition for Interactive Transcription in Australian Aboriginal Communities
Éric Le Ferrand · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Speech recognition is a field of Natural Language Processing that involves auto-matically recognising speech data and converting it into text, such as an orthographictranscription or a sequence of phones or other kinds of labels. These technologies arepredominantly used in Western societies and serve specific purposes, such as voicedictation, smart home devices, and automatic subtitling, for dominant languages suchas English, French, and Mandarin. Low-resource languages refer to languages forwhich available data is too scarce to solve a specific task.Indigenous communities worldwide have a history of Western scholars visitingtheir lands for research purposes. These visits generally have two goals: the firstis the collection and analysis of Indigenous data to help Western scholars gain abetter understanding of Indigenous contexts, and the second is the deployment ofWestern methods to address local issues. These methods are part of a long tradition ofcolonialist practices in which Westerners extract Indigenous resources without localbenefit or disregard Indigenous bodies of knowledge and ways of knowing.Recent calls have encouraged scholars to decolonise their research practices,which involves following a set of principles to enable more ethical research. Inresponse to this call, research projects need to benefit the community and respectits social, cultural and political components. These principles have been translatedinto the adoption of Indigenous research methods, negotiations with the communityregarding the outcome of each project, and working with the community to buildself-determination.The research around speech recognition has mainly focused on generalisinggeneric computational methods across different languages, without adequately ad-dressing ongoing colonial practices. In this thesis, we begin with the languagedocumentation agenda and explore how speech recognition technologies can fit intoIndigenous communities, following decolonising principles. Specifically, we focuson a single Indigenous Australian community and explore how to create a speechrecognition system that takes into account the constraints of this context. We alsoexamine how to use this system in a cross-cultural activity and shape it to respect thecommunity’s desired outcomes in terms of language work.