Audio wordlists as a basis for a lexicographic database (using Evenki, Selkup, and Ket data)

Abstract

The article presents a possible way of developing a lexicographic database from audio-recorded wordlists collected in the field as part of the documentation of three endangered Siberian languages: Selkup, Ket, and Evenki. During more than two decades of fieldwork (2001-2025), we have accumulated extensive audio materials, covering virtually all current settlement areas of these language communities. At present, all three languages are critically endangered, language transmission from parents to children has stopped in all Ket local communities and in the majority of Selkup and Evenki local communities, so it is essential to make the collected materials publicly available for linguistic research, language teaching, and revitalization efforts. However, processing such large volumes of field recordings poses significant technical challenges. We consider various initiatives on publishing field dictionaries, including the Dictionaria project and various language archive projects. We then describe our own digital dictionary platform for Evenki. The publication of field-recorded wordlists consists of the following steps: segmentation, transcription, and database creation. The first step is segmentation, i.e. the separation of the native speaker’s words from the linguist’s words. Segmentation was done manually using Audacity software. Then the audio files were transcribed. The article describes the automatic approach to the transcription using a fine-tuned MMS (Massively Multilingual Speech) model. For Evenki, we trained the model on approximately 8 hours of corpus data, having reduced the Character Error Rate (CER) from 0.46 to 0.2 on test data. This level of quality is by no means high, but it proves acceptable for indexing purposes when audio files remain accessible for verification. After that, we describe the web resource we developed using PostgreSQL, Django, and React. The web resource provides search functionality across all recorded wordlists with filtering by dialect, location, and recording date, as well as by the Russian stimulus words and the automatic transcription. Users can also export data and listen to audio files directly. At the end of the article, we demonstrate two case studies, which became possible due to database development: mapping the use of various terms for ‘bear’ and tracing a stem in a dialect where it had not been documented before. Contribution of the authors: O. Kazakevich: writing - original draft, writing - review & editing, supervision. E. Klyachko: writing - original draft, illustrations. The authors declare no conflicts of interests.