Language is the most natural mode of communication for a human being. It is intuitive and extremely versatile, allowing both clear instructions as well as subtle nuances of intention and opinion. The Spoken Language Systems group (LSV) of Saarland University works on unlocking these complexities. To advance the state of the art in meaningful ways, LSV pursues theoretical, fundamental and applied research as well as industry collaborations. Our research areas include

  • dialogue systems
  • images & language
  • knowledge extraction
  • hate speech detection
  • language modeling
  • low resource learning
  • privacy-preserving machine learning
  • sentiment analysis
  • speech recognition
  • question answering

Awards & Benchmarks

The LSV team repeatedly scored high in numerous international shared tasks competitions including information extraction, knowledge base population, sentiment and opinion analysis, information extraction from unstructured data. The team was also awarded the German Government Price `Deutschland – Land der Ideen’ for its innovative research and achievements in the area of Automatic Speech Recognition.

Current Projects

CRC Information Density and Linguistic Encoding

Language provides not only the expressiveness needed to communicate, but also offers speakers a multitude of choices regarding how they may encode their messages – from the choice of words, structuring of syntactic elements, and arranging sentences in discourse. The SFB addresses the hypothesis that language variation and language use can be better understood in terms of the goal of speakers to modulate the amount of information conveyed in an utterance. While previous efforts have sought to understand language systems and their use in terms of complexity, the definition of this notion is often imprecise and specific to particular linguistic levels. Recently, however, there is evidence that the ease of processing linguistic material is correlated with its contextually determined predictability. This has lead to the hypothesis that complexity may be appropriately indexed by Shannon’s notion of information, referred to in recent linguistic work as surprisal. The aim of the SFB is thus to investigate the extent to which notions of surprisal and the optimal distribution of information offer a unifying explanation of observed patterns of variation in language use within and across linguistic levels, and in a range of communicative settings.

More information


The H2020 project COMPRISE defines a fully private-by-design methodology and tools that will reduce the cost and increase the inclusiveness of voice interaction technology through research advances on privacy-driven data transformations, personalized learning, automatic labeling, and integrated translation. The vision is to filter out privacy-threatening portions of the user data before storing them in the cloud.

More information


The project aims at designing a framework for extracting evidence and actionable intelligence from large amount of noisy multilingual multimodal data based on advanced speech and language technologies (SLTs), visual analysis (VA) and network analysis (NA). The overall project goal is to achieve a significant improvement in identification of events, entities and relations, and to design a new generation of probabilistic and neural networks based tools interfacing SLT, VA and NA technologies.

More information


ATCO2 will deliver a platform to collect, store, process and share voice communications from real world air-traffic control data, exploiting deep learning methods. The planned machine learning solutions are enabling technologies for air-traffic control. To achieve robust and high speech recognition performance, large amount of data will be collected. The project aims at accessing data from certified ADS-B datalinks aligned with a surveillance technology, and directly from air-traffic controllers supplied by air navigation service providers.

More information