Musical Instrument Classification Using Deep Learning CNN Models

Musical instrument recognition is the act of counting on machine learning or the application of signal processing to isolate and classify various musical instruments in an audio track. Such capability makes it possible to better analyze musical pieces, which may be helpful whenever features need to be transcribed, songs recommended, or when they need to be composed automatically. In this work, we put forward a CNN based model for musical instruments classification using the FreeSound dataset. The model performs Mel-spectrograms on extracted audio inputs and make use of a highly effective data augmentation methodology known as CutMix. By using hyperparameter tuning and pruning approaches, the model configuration and structure is improved thereby enabling its effectiveness. Further, Class Activation Maps and an Ablation Study are employed to analyse interpretability and uncover the strength of the technique in instrument recognition. The accuracy was at 85%; the use of a richer database and convolutional neural networks may lift the accuracy of the classification. Using this algorithm means its application in a wider context of sound classification is possible and at some point, the algorithm can learn beyond a human's capability to differentiate between similar sounds (for example, distinguishing between viola and violin, which humans cannot do). Real world applications of musical instrument recognition are numerous. It can make the sorting and searching of enormous musical libraries possible with great efficiency, which would be valuable to streaming services since these offer their users relevant tune recommendations.