Detection of genre-specific musical instruments: The case of the mellotron

When facing the problem of organizing, categorizing, browsing and retrieving data from large music collections, musical instruments play a predominant role, as they define the timbral qualities in any piece of music. Recent technological developments in digital audio have made possible to automatize these tasks. Specific instruments can also be directly related to concrete musical genres, which increases the possible applications of such systems. This document addresses the problem of detection of musical instruments in polyphonic audio, exemplifying this specific task by analyzing the mellotron, a vintage sampler used in popular music. The mellotron presents interesting technical and perceptual qualities, which make it ideal for the study of timbre descriptors in the context of automatic classification in polyphonic audio. For accomplishing this task a novel methodology is presented, based on the idea that it is possible to train classifiers with audio descriptors (temporally integrated from the raw feature values extracted from polyphonic audio data) using extensive datasets. A series of experiments were designed in order to gather information about the specific descriptors that could help accomplish the detection and classification tasks, by employing custombuilt datasets classified according to instrumentation features. Several machine learning techniques are tested and evaluated according to the effectiveness of the system, that is, performance based on the accomplishment of the objectives by selecting different measures. The results obtained were relevant for the tasks proposed, with values far above chance in most cases, which indicates some statistical significance for assuring that the models tested are indeed recognizing the presence of the mellotron in a polyphonic context. The evidence shows that the methodology used proves to be effective for solving the task.