Fashioning Local Designs from Generic Speech Technologies in an Australian Aboriginal Community

Proceedings of the 29th International Conference on Computational Linguistics, pages 4274–4285 October 12–17, 2022. 4274 Fashioning Local Designs from Generic Speech Technologies in an Australian Aboriginal Community Éric Le Ferrand,1,2 Steven Bird,1 and Laurent Besacier2 1Northern Institute, Charles Darwin University, Australia 2Laboratoire Informatique de Grenoble, Université Grenoble Alpes, France Abstract An increasing number of papers have been addressing issues related to low-resource languages and the transcription bottleneck paradigm. After several years spent in North- ern Australia, where some of the strongest Aboriginal languages are spoken, we could ob- serve a gap between the motivations depicted in research contributions in this space and the Northern Australian context. In this paper, we address this gap in research by exploring the potential of speech recognition in an Aborig- inal community. We describe our work from training a spoken term detection system to its implementation in an activity with Aboriginal participants. We report here on one side how speech recognition technologies can find their place in an Aboriginal context and, on the other, methodological paths that allowed us to reach better comprehension and engagement from Aboriginal participants. 1 Introduction A consistent theme in recent NLP research has been doing more with less (Wiesner et al., 2022; Gao et al., 2021; Baevski et al., 2021; Schneider et al., 2019; Menon et al., 2019). It is popular to describe new pipelines to solve a wide range of tasks for under-resourced languages (Godard et al., 2018; Anastasopoulos et al., 2018; Settle et al., 2017; Mitra et al., 2016; Lane and Bird, 2019). However, the motivations behind the design of a computational method are not systematically well justified according to the needs of the target speech communities. The category of under-resourced languages en- compasses a wide range of contexts, not simply in terms of the quantity of data available but also in terms of local speech communities’ sociolinguistic and political situation (Bird, 2022). Often, the fo- cus has been to generalise a given method across languages, where the proposed system is at the core of the argument instead of the benefits that it could have for the speakers. We could ask whether the same language technology would be equally applicable to Marathi, spoken by millions in a ma- jor metropolis, and Miriwoong, with only a few elderly speakers in a remote Australian Aboriginal community (cf. Kuhn, 2022). Universal solutions dominate NLP: research and results are often provided without taking into ac- count the global situation of the languages involved or the views of the speech communities about the preservation of their language. Instead, it is com- mon to assert that an improvement in Word Error Rate yielded by a given speech recognition sys- tem is the answer to the transcription bottleneck and, therefore, the problem of scaling up language documentation (van Esch et al., 2019; Foley et al., 2018). Most of the world’s languages are primarily oral (Ong, 1982; Walsh and Yallop, 1993). Writing is often not a priority, and very few people are skilled in transcribing their …

Content truncated to 500 words