Towards automatic annotation of prosodic prominence levels in Austrian German

The creation of prosodic annotations is one of the most diffi-cult and time-consuming aspects of creating a speech database. Generally, only the speech signal and manually created tran-scriptions are available in an early resource development stage.This paper presents a tool for annotating prosodic prominenceat the word level, using exclusively acoustic features (96 f0-,intensity- and durational features). The best performance forseparating prominent from non-prominent words in Austrianread speech was reached with a decision tree with the abso-lute word duration as the only feature. For distinguishing moreprominence levels, a good performance was reached with a ran-dom forest model, similar to the best inter-annotator agreement.Furthermore, we analyzed in detail the feature ranking of therandom forest to give us insights into the relative importanceof the features contributing to prominence in Austrian German:Word duration>f0 range, RMS range. The specific findingsof this study will mainly be relevant for speech scientists andprosody researchers interested in German. Our methodologi-cal approach of analyzing prosodic prominence from a purelyacoustic perspective at the word-level will also be interestingfor researchers focusing on prosody in other languages.