Comparison of features in musical instrument identification using artificial neural networks

This paper examines the use of a number of auditory features in identifying musical instruments. The Temporal Envelope, Centroid, Melfrequency Cepstral Coefficients (MFCCs), Inharmonicity, Spectral Irregularity and Number of Spectral Peaks are all examined. By using these features to train a Multi-Layered Perceptron (MLP), it is determined that the MFCCs are the most efficient of these features in musical instrument identification. The Inharmonicity, Spectral Irregularity and Number of Spectral Peaks offered no benefit to the classifier. Of the instruments studied, the piano was most accurately classified and the violin was the least accurately classified instrument.