Detecção de Onsets em notas de músicas instrumentais de piano utilizando representação Pitch e aprendizado de máquina

Abstract

The analysis of music signals and the extraction of musically meaningful information to build musical applications are part of the research field called music information retrieval (MIR), within which the task of automatic onset detection is inserted. The detection of onsets in musical signals consists in detecting the time instants of the beginning of the musical events contained in the musical signal, and this task, generally, serves as a base for building applications like automatic music transcription of one or more musical instruments, audio-score alignment, time estimation of music, among others. This dissertation presents a system for automatic onset detection in piano music signals using machine learning. In the proposed framework, the time-frequency representation pitch is used and the classifiers investigated are support vector machine (SVM), gradient boosting, and one-dimensional convolutional neural network (1D CNN). The experiments results made with the databases BS1 and MAESTRO show that, in the first approach, the SVM had superior performance compared with gradient boosting, while in the second approach, the sensibility metric was higher when using the pitch features instead of the spectrogram features on the BS1 database.