Quantifying the relevance of locally extracted information for musical instrument recognition from entire pieces of music

In this work we study the problem of computational musical instrument recognition from entire pieces of music. In particular, we present and evaluate 4 different methods to select, from an unknown piece of music, relevant excerpts in terms of instrumentation, on top of which instrument recognition techniques are applied to infer the labels. Since the desired information is assumed to be redundant (we may extract a few labels from a thousands of audio frames) we examine the recognition performance, the amount of data used for processing, and their possible correlation. Experimental results on a collection of Western music pieces reveal state-of-the-art performance in instrument recognition together with a great reduction of the input data. However, we also observe a performance ceiling with the currently applied instrument recognition method.