Polyphonic music information retrieval based on multi-label cascade classification system

Abstract

With the fast booming of online music repositories, there are increasing needs for content-based Automatic Indexing to help users find their favorite music objects. Melody matching based on pitch detection technology has drawn much attention and many MIR systems have been developed to fulfill this task. However music instrument recognition remains an unsolved problem in the domain. Numerous approaches to acoustic feature extraction have already been proposed. Unfortunately, none of those monophonic (one distinct instrument) timbre estimation algorithms can be successfully applied to polyphonic (multiple distinct instruments) sounds, which occur more often in the real music world. This has stimulated the research on multi-labeled instrument classification and new features development for content-based automatic music information retrieval. The original audio signals are a large volume of unstructured sequential values, which are not suitable for traditional data mining algorithms, while the higher level data representative of acoustical features are sometimes not sufficient for instrument recognition in polyphonic sounds. We propose a multi-labeled classification system to estimate multiple timbre information from the polyphonic sound according to a similarity measure based on both feature vectors and spectrum envelope. In order to achieve a higher estimation rate, we introduced the hierarchical structured classification model under the inspiration of the human perceptual process. This cascade classification system would first estimate the higher level decision attribute, which stands for the musical instrument family. Then further estimation would be done within that specific family range. This could be applied with different kind of features according to the specific characteristics of instruments in this family. Experiments showed better performance of cascade system than the flattened classifiers. © 2009 Springer-Verlag Berlin Heidelberg.