A Novel Convolutional Neural Network Model for Musical Instruments’ Classification: A Deep Signal Processing Approach

Stress management is a challenge in this modern world, and several methods are being practiced such as diet, exercise, sleep, meditation and relaxation. One of the relaxation methods is the Music Therapy, where one listens to Vocal music, Instrumental music, Hindustani classical music, and Western music etc. In instrument music, various types of instruments are used and these instruments vary from country to country, which produce different sounds. Based on the sounds, humans recognize these instruments with certain efficacy that varies with the expertise. In this paper, we propose the classification of 14 categories of musical instruments, namely, bells, cello, clarinet, crotales, double Bass, flute, piano, saxophone, trombone, trumpet, vibraphone, viola, violin, and xylophone etc. based on the sound signal. A total of 8,365 spectrograms, across 14 classes, are used. We have deployed five pre-trained CNN models such as Efficientnet, Googlenet, ResNet, Squeezenet, and Mobilenet for transfer learning. Since the classification accuracies of these pre-trained models are not up to the expectations, hence, a custom designed architecture is proposed, which outperforms the pre-trained models giving 99.81 % classification accuracy. Hyper-parameters are varied during the experimentation and the results are compared with the state-of-the-art methods. The work is helpful for the practicing psychiatrists, in knowing what types of sounds manage, which kinds of stresses.