Automatic extraction of music descriptors from acoustic signals

High-Level music descriptors are key ingredients for music information retrieval systems. Although there is a long tradition in extracting information from acoustic signals, the field of music information extraction is largely heuristic in nature. We present here a heuristicbased generic approach for extracting automatically high-level music descriptors from acoustic signals. This approach is based on Genetic Programming, used to build relevant features as functions of mathematical and signal processing operators. The search of relevant features is guided by specialized heuristics that embody knowledge about the signal processing functions built by the system. Signal processing patterns are used in order to control the general processing methods. In addition, rewriting rules are introduced to simplify overly complex expressions, and a caching system further reduces the computing cost of each cycle. Finally, the features build by the system are combined into an optimized machine learning descriptor model, and an executable program is generated to compute the model on any audio signal. In this paper, we describe the overall system and compare its results against traditional approaches in musical feature extraction à la Mpeg7.