Designing a Training Set for Musical Instruments Identification

Abstract

This paper presents research on one of the most challenging branches of music information retrieval – musical instruments identification. Millions of songs are available online, so recognizing instruments and tagging them by a human being is nearly impossible. Therefore, it is crucial to develop methods that can automatically assign the instrument to the given sound sample. Unfortunately, the number of well-prepared datasets for training such algorithms is very limited. Here, a series of experiments have been carried out to examine how the mentioned methods’ training data should be composed. The tests were focused on assessing the decision confidence, the impact of sound characteristics (different dynamics and articulation), the influence of training data volume, and the impact of data type (real instruments and digitally created sound samples). The outcomes of the tests described in the paper can help make new training datasets and boost research on accurate classifying instruments that are audible in the given recordings.