Using Audio Loops for Instrument Family Recognition in Machine Learning Tasks

Abstract

This paper introduces an instrument recognition approach with the aid of audio loops. The aim is to show a basic instrument recognition recipe for music technology researchers by investigating whether the DAW-based audio loops can be an alternative to researched-based available libraries such as McGill University master samples, UIOWA samples, IRMAS audio libraries. For that purpose, audio loops from Apple Jam Pack were preferred to create instrument classes (Families). The loops were arranged according to their related instrument classes. The class names are Bass, Drums and Percussions, Guitars, Keyboards, Strings, Synthesizers, and Winds. After the extraction of temporal and spectral audio features from those classes, a 5736x105 dimensional dataset emerged. Then this dataset was examined with 19 different supervised machine learning algorithms. The SVM Cubic classification algorithm provided the best accuracy (90.2%). The result shows that the audio loops with mid-term feature extraction can be used for instrument recognition tasks.