Learning feature extraction for learning from audio data

Today, large collections of digital music plays are available. These audio data are time series which need to be indexed and classified for diverse applications. Indexing and classification differs from time series analysis, in that it generalises several series, whereas time series analysis handles just one series a time. The classification of audio data cannot use similarity measures defined on the raw data, e.g. using time warping, or generalise the shape of the series. The appropriate similarity or generalisation for audio data requires feature extraction before classification can successfully be applied to the transformed data. Methods for extracting features that allow to classify audio data have been developed. However, the development of appropriate feature extraction methods is a tedious effort, particularly because every new classification task requires to tailor the feature set anew. Hence, we consider the construction of feature extraction methods from elementary operators itself a first learning step. We use a genetic programming approach. After the feature extraction, a second process learns a classifier from the transformed data. The practical use of the methods is shown by two types of experiments: classification of genres and classification according to user preferences