Musical Instrument Identification and Tone Detection Using Feature Learning

Abstract

In music information retrieval (MIR) field, most of the tasks have been heavily relying on the hand-crafted features which are highly useful to measure and quantify the certain target characteristics of the sound such as pitch, roughness, and brightness. However, there are an increasing number of attempts on applying feature learning technique, which has shown superior performance across the research fields, in MIR tasks especially when the goal is the identification of the sound. The aim of this thesis is to advance the state-of-the-art in musical instrument identification and tone detection tasks with feature learning approaches, which can be used for various music-related applications, including but not limited to, music searching, browsing, recommendation, and education. We utilize sparse feature learning and convolutional neural network to learn a feature from input data, and propose the network architecture and data processing framework that is suitable for music signal. We present experimental results on MIR tasks such as fingering detection of overblown flute sound, instrument identification in monophonic sound, and predominant instrument recognition in a real-world polyphonic music. In addition, we conducted an extensive experiment to find the optimal data processing techniques and parameters such as input frame sampling, frequency scaling, activation pooling, window/hop size, output aggregation method, and network training hyperparameters.