Today, large audio collections are stored at computers. Their organization can be supported by machine learning. This demands a more abstract representation than is the time series of audio values. We have developed a unifying framework which decomposes the complex extraction methods into their building blocks. This allows us to move beyond the manual composition of feature extraction methods. Several of the well-known features as well as some new ones have been composed automatically by a genetic learning algorithm. While this has delivered good classifications it needs long training times. Hence, we additionally follow a meta-learning approach. We have developed a method of feature transfer which exploits the similarity of learning tasks to retrieve similar feature extractions. This method achieves almost optimal accuracies while it is very efficient. Nemoz, an intelligent media management system, incorporates adaptive feature extraction and feature transfer which allows for personalized services in peer-to-peer settings.