Accessing spoken interaction through dialogue processing

Written language is one of our primary means for documenting our lives, achievements, and environment. Our capabilities to record, store and retrieve audio, still pictures, and video are undergoing a revolution and may support, supplement or even replace written documentation. This technology enables us to record information that would otherwise be lost, lower the cost of documentation and enhance high-quality documents with original audiovisual material. The indexing of the audio material is the key technology to realize those benefits. This work presents effective alternatives to keyword based indices which restrict the search space and may in part be calculated with very limited resources. Indexing speech documents can be done at a various levels: Stylistically a document belongs to a certain database which can be determined automatically with high accuracy using very simple features. The resulting factor in search space reduction is in the order of 4-10 while topic classification yielded a factor of 18 in a news domain. Since documents can be very long they need to be segmented into topical regions. A new probabilistic segmentation framework as well as new features (speaker initiative and style) prove to be very effective compared to traditional keyword based methods. At the topical segment level activities (storytelling, discussing, planning, . . .) can be detected using a machine learning approach with limited accuracy; however even human annotators do not annotate them very reliably. A maximum search space reduction factor of 6 is theoretically possible on the databases used. A topical classification of these regions has been attempted on one database, the detection accuracy for that index, however, was very low. At the utterance level dialogue acts such as statements, questions, backchannels (aha, yeah, . . .), etc. are being recognized using a novel discriminatively trained HMM procedure. The procedure can be extended to recognize short sequences such as question/answer pairs, so called dialogue games. Dialog acts and games are useful for building classifiers for speaking style. Similarily a user may remember a certain dialog act sequence and may search for it in a graphical representation. In a study with very pessimistic assumptions users are able to pick one out of four similar and equiprobable meetings correctly with an accuracy ≈ 43% using graphical activity information. Dialogue acts may be useful in this situation as well but the sample size did not allow to draw final conclusions. However the user study fails to show any effect for detailed basic features such as formality or speaker identity.