||To build highly accurate and robust machine learning algorithms practitioners require data in high quality, quantity and diversity. Available time series data sets often lack in at least one of these attributes. In cases where collecting more data is not possible or too expensive, data-generating methods help to extend existing data. Generation methods are challenged to add diversity to existing data while providing control to the user over what type of data is generated. Modern methods only address one of these challenges. In this thesis we propose a novel generation algorithm that relies on characteristics of time series to enable control over the generation process. We combine classic interpretable features with unsupervised representation learning by modern neural network architectures. Further we propose a measure and visualization for diversity in time series data sets. We show that our approach can create a controlled set of time series as well as adding diversity by recombining characteristics across available instances.