Automatic analyzes and descriptions of soundscapes, which are qualitatively similar to human descriptions, are very impor tant to objectify soundscape research and to develop convincing tools for soundscape analysis. One possible way to develop such an automatic soundscape analysis system is to start with an annotation system for unrestricted real-world sounds and enrich this system with more and more automated functions. We will demonstrate an annotation system that automatically selects coherent time-frequency regions of similar spectro-temporal proper ties and similar acoustic ¿textures¿. The artificial system will learn through the interaction with a user whose task it is to assign names to regions. According to prototype theory of categorization, naming would structure the acoustic signals into meaningful relevant categories integrating textures and properties along family resemblance. After a low number of examples the system is able to associate an annotation with a time-frequency region. In cognitive terminology, the system will be able to use the emerging categories for evaluating new signals through top-down processing and membership decisions. We propose that systems like this can be used to bootstrap a more general, and eventually automatic, soundscape analysis system that complies with the aim stated above.