Multitimbral musical instrument classification

The automatic identification of musical instrument timbres occurring in a recording of music has many applications, including music search by timbre, music recommender systems and transcribers. A major difficulty is that most music is multitimbral, making it difficult to identify the individual timbres present. One approach is to classify music based on specific groups of musical instruments. In this paper we report on our experiments that classify musical instrument timbres based on specific groups that are often found in commercial recordings. Classification using the K-Nearest Neighbour classifier, with audio features such as Mel Frequency Cepstral Coefficients, on a set of 160 samples from commercial recordings, gave an accuracy of 80%. Some of the difficulties arose from distinguishing similar instrument groups, such as those only differing by the inclusion or exclusion of a voice. However, when these were examined in isolation, greater accuracy was achieved, suggesting that a hierarchical approach may be helpful.