Automatic Speech Recognition for Supporting Endangered Language Documentation

Generating accurate word-level transcripts of recorded speech for language documentation is difficult and time-consuming, even for skilled speakers of the target language. Automatic speech recognition (ASR) has the potential to streamline transcription efforts for endangered language documentation, but the practical utility of ASR for this purpose has not been fully explored. In this paper, we present results of a study in which both linguists and community members, with varying levels of language proficiency, transcribe audio recordings of an endangered language under timed conditions with and without the assistance of ASR. We find that both time-to-transcribe and transcription error rates are significantly reduced when correcting ASR for language learners of all levels. Despite these improvements, most community members in our study express a preference for unassisted transcription, highlighting the need for developers to directly engage with stakeholders when designing and deploying technologies for supporting language documentation.