ASR for Documenting Acutely Under-Resourced Indigenous Languages

ASR for Documenting Acutely Under-Resourced Indigenous Languages Robbie Jimerson, Emily Prud’hommeaux Rochester Institute of Technology Rochester, New York, USA rcj2772,emilypx@rit.edu Abstract Despite its potential utility for facilitating the transcription of speech recordings, automatic speech recognition (ASR) has not been widely explored as a tool for documenting endangered languages. One obstacle to adopting ASR for this purpose is that the amount of data needed to build a reliable ASR system far exceeds what would typically be available in an endangered language. Languages with highly complex morphology present further data sparsity challenges. In this paper, we present a working ASR system for Seneca, an endangered indigenous language of North America, as a case study for the development of ASR for acutely low-resource languages in need of linguistic documentation. We explore methods of leveraging linguistic knowledge to improve the ASR language models for a polysynthetic language with few high-quality audio and text resources, and we propose a tool for using ASR output to bootstrap new data to iteratively improve the acoustic model. This work serves as a proof-of-concept for speech researchers interested helping field linguists and indigenous language community members engaged in the documentation and revitalization of endangered languages. Keywords: low-resource languages, automatic speech recognition, indigenous languages 1. Introduction By the end of this century, it is estimated that at least half and as many as 90% of the world’s nearly seven thousand languages will be extinct (Krauss, 1992; Crystal, 2000). With each language that is lost, we lose insight not only into the culture of the the people who spoke that language, but also into the characteristics of that language that can shed light on the underlying structure of human language. Some communities on the verge of losing their language are engaged in preservation efforts, and many linguists carry out field work with native speakers to document endangered languages. Automatic speech recognition (ASR) has the potential to serve as a useful tool in these preservation and documentation efforts, but building models for these languages presents numerous challenges. One particular challenge is a dearth of data, specifically, transcribed and labeled audio data to train the acoustic model and large amounts of text to train the language model. Languages that lack extensive data are known as under-resourced or low-resource languages, and all but a handful of the world’s languages fall into this category. In fact, the set of languages considered to be low-resource for the purposes of ASR research includes many very widely spoken languages, including Bengali (spoken na- tively by 200 million people) and Vietnamese (with more native speakers than French) (Harper, 2014). Although researchers do not currently have access to large amounts of labeled data in these languages, it would be relatively easy to acquire more data with small investments of time and money to train speakers of the language to collect and label data. Languages like Vietnamese and Bengali, which for po- litical and economic reasons happen to have few ASR resources, stand in contrast to what we will call acutely …

Content truncated to 500 words