Detection and clustering of musical audio parts using Fisher linear semi-discriminant analysis

We present a method aiming at facilitating musical audio summarization by organizing the signal into a set of possibly recurring parts, such that inclusion of an expert from each part would be adequate to compactly summarize the whole audio signal. Crucial to the success of the grouping segments into parts is the underlying distance metric, which depends on the feature space and should provide distances that are low for segments of the same audio part and high for segments of different audio parts. Starting with a general purpose audio feature space, we use the information from the sequential structure of audio signals, in order to estimate in a completely unsupervised way a Fischer subspace with discriminant characteristics for the particular audio signal. The derived feature space is used in a segmentation-clustering system based on fuzzy clustering, HMM and k-NN probability estimation. The experimental results show an almost 10% performance gain when adopting the Fisher subspace with respect to using the original feature space.