Prosody in Speech Recognition and Understanding
October 22-24, 2001
This paper provides a prosodic analysis of so-called Off-Talk in spoken German in human-computer interaction. Off-Talk consists of user utterances, which are not directed to the automatic speech processing system. These utterances have to be mastered automatically as far as possible. The data collection in the SmartKom project is described and problems with the consistent annotation of Off-Talk are discussed. Different forms of Off-Talk are distinguished and compared prosodically with each other as well as with other speech, which is directed to the system. The analysis includes various perceptual and acoustic prosodic features. There are clear differences in the distribution of prominent accents and phrase boundaries found. F0 range turned out to be a further relevant feature. In future, a refined definition of Off-Talk has to be applied, which fulfills our requirements for consistent and efficient annotation.
Full Paper (PDF) Full Paper (Zipped Postscript)
Bibliographic reference. Siepmann, Rolf / Batliner, Anton / Oppermann, Daniela (2001): "Using prosodic features to characterize Off-Talk in human-computer interaction", In Prosody-2001, paper 27.