Abstract
The paper presents a spoken corpus of the endangered Torlak dialect from the Timok area of Southeast Serbia. This dialect expresses a great deal of variation in the use of non-standard features under the influence of standard Serbian (SSr).
Accounting for this variation, a specific methodology has been selected for collection, sampling, transcription and annotation. Between 2015 and 2017, semi-structured interviews were conducted in the field eliciting spontaneous speech in the form of long narratives about traditional culture and history.
The corpus comprises 500,697 tokens of semi-orthographic transcripts representing 80 h of recording from locations evenly distributed across the Timok area of the Torlak dialect zone, thus enabling a spatial contrastive analysis. The majority of speakers in the corpus are older people whose language represents the highly non-standard variety.
In order to allow for analysis of language change under the influence of SSr, the corpus includes a number of younger people whose speech is closer to SSr. Tools for automatic PoS annotation and lemmatization that were lacking were developed based on the existing resources for SSr.
For tagger training, a dialect sample of 27,000 manually verified tokens was merged with an existing training set for SSr.