Sound-aligned corpus of Udmurt dialectal texts

The paper describes an ongoing effort aiming at building a sound-aligned corpus of Udmurt spoken texts. The corpus currently consists of about 3.5 hours of recordings, collected during fieldwork trips between 2014 and 2016. The recordings represent three dialect groups of Udmurt (Northern, Central and Southern). The recordings were transcribed with the help of native speakers. All morphological peculiarities characteristic of spoken or dialectal Udmurt were faithfully reflected, however, the transcription was somewhat normalized in order to facilitate morphological annotation and cross-dialectal search. The pipeline of our project includes aligning the texts with the sound in ELAN and annotating them with a morphological analyzer developed for standard Udmurt. We use automatic annotation as a much less time-consuming alternative of manual glossing and explore the resulting quality and the downsides of such annotation. We are specifically investigating how much and what kind of change the standard analyzer requires in order to achieve sufficiently good annotation of spoken/dialectal texts. The corpus has a web interface where the users may execute search queries and listen to the audio. The online interface will be made publicly available in 2018.