Enabling Interactive Transcription in an Indigenous Community

Proceedings of the 28th International Conference on Computational Linguistics, pages 3422–3428 Barcelona, Spain (Online), December 8-13, 2020 3422 Enabling Interactive Transcription in an Indigenous Community ´Eric Le Ferrand,1,2 Steven Bird,1 and Laurent Besacier2 1Northern Institute, Charles Darwin University, Australia 2Laboratoire Informatique de Grenoble, Universit´e Grenobles Alpes, France Abstract We propose a novel transcription workflow which combines spoken term detection and human- in-the-loop, together with a pilot experiment. This work is grounded in an almost zero-resource scenario where only a few terms have so far been identified, involving two endangered languages. We show that in the early stages of transcription, when the available data is insufficient to train a robust ASR system, it is possible to take advantage of the transcription of a small number of isolated words in order to bootstrap the transcription of a speech collection. 1 Introduction In remote Aboriginal communities in Australia, many efforts are made to document traditional know- ledge including rock art, medicinal plants, and food practices. While it may be relatively straightforward to capture spoken content, transcription is time-consuming and has been described as a bottleneck (Brinckmann, 2009). Transcribing is often seen as an obligatory step, to facilitate access to audio. Efforts have been made to speed up this process using speech recognition systems, but the amount of data available in Indigenous language contexts is usually too limited for such methods to be effective. Recent research has shown the efficacy of spoken term detection methods when data are scarce (Menon et al., 2018a; Menon et al., 2018b). Taking advantage of the transcription of a few words would allow us to propagate it through the speech collection and thus assist language workers in their regular transcription work. So-called “sparse transcription” would be also a way to navigate a speech collection and allow us to be selective about what needs to be transcribed (Bird, 2020). Several tools exist for manual transcription, such as Elan and Praat (Wittenburg et al., 2006; Boersma and Weenink, 1996). However such transcriptions are often made in isolation from the speech community (First Languages Australia, 2014), and so we miss out on the opportunity to take advantage of the interests and skills of local people to shape and carry out the transcription work. We present a fieldwork pipeline which combines automatic speech processing and human expertise to support speech transcription in almost-zero resource settings. After giving the background, we detail the workflow and propose a pilot experiment on two very low-resource corpora. 2 Background Existing approaches to automatic transcription of endangered languages involve methods that have been developed for automatic speech recognition. While a few hours of transcribed speech can be enough to train single-speaker models (Gupta and Boulianne, 2020b), speaker-independent models require a large amount of training data to produce useful transcriptions (Gupta and Boulianne, 2020a; Foley et al., 2018). Moreover, they draw language workers and speakers into the time-consuming task of exhaustive transcription, forcing them to transcribe densely, including passages that may be difficult or impossible given the early state …

Content truncated to 500 words