Computational modeling of tone in language documentation: citation tones vs. running speech in Chindwin Khamti [Paper]

This paper examines how contextual variables can explain the significant gap in performance for unsupervised modeling of tones in Tai Khamti [ISO 639-3: kht] spoken in Myanmar. Two corpora were extracted from citation tones and tones in running speech in order to assess the utility and limitations of these methods. Taking native judgments as ground truth, current results show high precision on citation tones, between 0.93 and 1.0, in three of the four expected tonal categories, as well as recall 0.79-0.86 in all four. Tones in sentential contexts showed precision just 0.28-0.62, with recall between 0.21 and 0.63.