Sentence boundaries in text and pauses in speech: Correlation or confrontation?

The paper explores the interaction between sentence boundaries marked by annotators in transcriptions of Russian spontaneous speech and actual prosodic boundaries in the signal. The aim of the research is to investigate whether annotators’ prosodic competence allows them to correctly detect sentence boundaries in speech based on textual information only. We found that inter-annotator agreement for each sentence boundary identified in transcription was affected by both presence or absence of pause and pause duration. Mixed linear model showed that presence or absence of pause explain 13% of variance in boundary detection. Pause duration explained only 4% of variance in inter-annotator agreement with moderate correlation of r = 0.21. We argue that relatively small size of effect in this case may be due to the interaction of different pausing strategies typical for reading and spontaneous speech, ambiguity of sentence boundaries and individual differences in speech perception.