Evaluation of Detrended Fluctuation Analysis applied to audio signals from drum cymbals in different machine learning classification contexts

Drum cymbals are musical instruments whose acoustic complexity is built from the union of various factors, from their constituent material to the final touch in the manufacturing process. The use of machine learning techniques to classify cymbals has been investigated in the literature over the years. However, most of these studies use a simplified strategy based on extracting features from signals. With this in mind, this paper explores the effectiveness of applying the combined triggering and Detrended Fluctuation Analysis (DFA) pre-processing techniques compared to using a set of time-domain information retrieved from the signals. To carry out this task, two contexts for classifying cymbals were considered: classification (i) by their bronze alloys (3 classes) and (ii) by each cymbal individually (4 classes). In addition, five machine learning models were evaluated. The Karhunen-Loève transformation based Classifier (KLTbC) performed well in both the three-class and the four-class classification schemes, achieving high accuracy with small standard deviations. In addition, the computational experiments showed that a trigger was useful to identify the initial moment of signal, and that the DFA could be applied as a pre-processing step, reducing data with a smaller number of points than the original signals, indicating that this is a promising strategy for the problem.