Musical Instrument Identification Using Machine Learning

Abstract

Music is a beautiful form of expression that is universally understood. In fact, people across all cultural backgrounds, even those without extensive exposure to musical education, can easily recognize the unique sounds of different instruments. This paper explores the use of machine learning to identify various instruments. The scope of this project focuses on audio recordings with one instrument playing one pitch, and the specific instruments in this study were viola, piano, and ukulele. This paper analyzes ways to quantify the differences in sounds of instruments using the harmonic frequency content apparent in spectrograms. It proposes the use of a simple but efficient K-Nearest-Neighbors machine learning algorithm. The results achieved 80% accuracy; using a larger dataset and using convolutional neural networks could improve accuracy of classification. The machine learning algorithm can be applied to the broader world of sound classification, and it can eventually surpass a human’s ability at identifying sounds (e.g. telling apart viola from violin, which the average human cannot do). There are numerous real-world applications for musical instrument recognition. It can enable automatic sorting and searching of massive musical collections with ease, which could be useful to streaming platforms that provide personalized recommendations to their users.