One important feature of linguistic communication is that some parts of utterances are more prominent than others. Prominence as a perceptual feature of spoken language is influenced by many different linguistic variables, but it is not clear how these variables interact in perception and which variables are most important for determining prominence. We report results from a prosody transcription task which assessed how untrained German listeners are simultaneously affected by gradient signal-based factors such as pitch, intensity and duration, as well as discrete prosodic factors (pitch accent type and placement) and non-prosodic factors (semantic-syntactic, lexical). All 17 linguistic variables tested were reliably associated with listeners’ prominence judgments. We used random forests, a data mining algorithm, to uncover which variables are most important in determining the prominence judgments. This analysis showed that discrete prosodic variables relating to intonational phonology, specifically the type of pitch accent and its position, were most predictive of prominence. At the same time, prominence judgments were characterized by large individual differences. An exploratory cluster analysis suggests that some listeners pay more attention to pitch-related variables but less to semantic-syntactic and lexical variables, while others do the reverse. Our results paint a complex picture of prominence perception that is highly variable across listeners, but that assures robust communication of prominence information through the simultaneous perceptual cueing of many different linguistic variables.
Bibliographical noteAccepted May 12
- Random forests
- mixed models