Iracema: from note onset detection challenges towards an audio content analysis library for the empirical study of music performance

• Ano de defesa: 2021
• Autor(a) principal: Tairone Nunes Magalhães
• Orientador(a): Mauricio Alves Loureiro
• Banca de defesa: Jose Augusto Mannis, Hugo Bastos de Paula, Flávio Luiz Schiavoni, Sérgio Freire Garcia, Thiago de Almeida Magalhães Campolina, Davi Alves Mota
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Música
• País: Brasil
• Palavras-chave em Português: Empirical musicology; Note onset detection; Music performance; Music information retrieval; Machine learning
• Link de acesso: http://hdl.handle.net/1843/44588

The earliest empirical studies on music performance date back to the end of the nineteenth century, when the first mechanical devices capable of recording the actions of pianists on the instrument (key presses) were invented. Since then, many technologies that open up new possibilities for collecting data from musical performances have been invented or developed, including techniques for extracting information directly from audio recordings. These techniques, which have been driven by the fast-paced technological development in computer-related fields over the last decades, are the subject matter of this thesis. We introduce a new software library called Iracema, which contains techniques for extracting patterns of manipulation of timing, energy, and spectral content from monophonic audio recordings. In this endeavor, the clarinet is the instrument chosen for the baseline experiments and models, but most of the presented techniques should also work for other monophonic instruments. One of the most critical steps in studying musical performances is the detection of the note onsets because our perception of timing is strongly tied to this variable. We pay special attention to this topic, proposing an interactive web interface for the precise manual annotation of note onsets and conducting an experiment to assess the typical measurement error involved in this kind of task for clarinet recordings. We also propose an annotated dataset of solo clarinet recordings containing approximately 23 minutes of audio and a total of 3551 note onsets. Using this dataset, we train a convolutional neural network to generate a model for automatic note onset detection specifically on clarinet recordings and compare its results to other onset detection models. Finally, we discuss a study case using recordings of a clarinet excerpt by a few different clarinetists to demonstrate the use of the proposed library.