Exploiting cepstral coefficients and CNN for efficient musical instrument classification

Identification of musical instruments is a vital problem in the area of Music Information Retrieval. Automatic music classification provides the foundation for a variety of advanced AI applications in the musical domain, such as automatic multi-instrument classification and information extraction from polyphonic audio. In this work, we construct a Convolutional Neural Network (CNN) and train it to classify twenty different musical instrument classes. This work makes use of the Philharmonia dataset, which contains twenty classes of instruments belonging majorly to four instrument families: brass, woodwinds, percussion, and strings. We extract the mel-frequency cepstral coefficients (MFCCs) of the sounds and use them as the input representation for our model. Our work also exploits audio data augmentation techniques on a minority class of the highly imbalanced dataset. To ensure instrumentalist independence, instrument sounds belonging to 14 classes of instruments, that match those from the Philharmonia, are taken from the UIOWA MIS database. The model is trained and tested using this joint dataset. Using the CNN we obtain a new state-of-the-art accuracy.