Aprendizado de máquina para a classificação automática de pratos de bateria conforme a proporção de estanho presente em suas ligas de bronze

Abstract

The many factors that influence the sound of a cymbal, combined with the external aspects that modify its sound perception, make the study of the acoustics of these instruments more challenging. Research that focuses on the study of drum cymbals is found in the literature, with various approaches and methods of investigation. However, considering the context of machine learning, the application of this type of technique to the study of drum cymbals is still little explored and usually focus on identifying the categories of cymbals and the way they were played. Thus, the use of machine learning-based computational tools to investigate the acoustic elements of these instruments becomes a topic of study with potential for further development. Inspired by this, this paper proposes classifying drum cymbals according to their constituent materials, since the metal alloy assumes a significant share of responsibility in their acoustics. To this end, a procedure for the standardized capture of the sounds of these instruments was developed and allowed to build a database consisting of 276 audios, referring to 4 cymbals. Different sets of attributes, selection techniques and classifiers were evaluated in order to propose an efficient computational model to solve this problem. As a result, an average accuracy of 99.75% was achieved when applying logistic regression on a subset of 30 attributes formed by 17 temporal features and 13 Mel-Frequency Cepstral Coefficients, selected by a technique based on the analysis of the correlation threshold between variables. On the other hand, when prioritizing the dimensionality of the selected subset, the use of a selection technique based on LASSO regression followed by correlation analysis, provides to identify a subset of 5 attributes that promote an average accuracy of 99.37%, also using logistic regression as a classifier.