Sebastian Möller, German Research Center forr Artificial Intelligence (DFKI) Berlin and Quality and Usability Lab, TU Berlin
In this overview talk, I would like to shed light on recent activities our teams have performed in speech and language technologies, and in human-machine interaction in a larger sense. We start from core technologies for extracting information from text, which we use to detect disinformation, which we consider as a hybrid task, and analyze how human evaluators can be supported best using machine classifiers. This requires explanations of model behavior, which we address via textual and dialogic explainability. Spoken and written dialog and its evaluation is also another topic of our research. On the speech side, we continue with our dimension-based speech quality assessment framework, which is also useful for disentangling speaker characteristics, used e.g. in speaker anonymization. We also work on the effect of listener characteristics on speech quality, as well on evaluating speech quality in a conversational scenario. Finally, I briefly summarize some work on XR interfaces, used e.g. for conveying sign language. The models and tools we use in our research are commonly shared with the community, both on a national and european level.
