A novel tool for capturing conceptualized audio annotations

For each supervised classification task some sort of ground truth data is needed in order to train the data models or classifiers and to evaluate the obtained result. Although there are a number of such data sets publically available for mainstream audio and music classification tasks, most often one will end up annotating new content by oneself when a novel or a specialized classifier needs to be developed. Though often necessary, the gathering of manually annotated metadata is a time-consuming and expensive exercise. Moreover, such metadata need to be structured in a proper way and assigned to the respective audio excerpts in order to be able to automatically process them. In this paper we present a novel software tool that facilitates the gathering of conceptualized annotations for any kind of audio content. The tool can be configured using arbitrary annotation schemas, which makes it flexible for multiple application fields. It furthermore provides automated audio s egmentation which helps to intuitively navigate through different parts of the audio file during the annotation process and select the right segment. The tool was originally developed to assist musicologists in collecting detailed metadata for global music contents, but it turned out to be more widely applicable, e.g. for annotating audiobooks or podcasts.