Resumo:
Extracting information about a guitar effects pedal being applied to an audio is oneof the tasks available for the field of musical content analysis. This study investigates the viability of the usage of two time-frequency representations of an audio, the HCQT and the wavelet scalogram to build a classifier based on a convolutional neural network that’s able to identify what effect is being applied in a pre-recorded audio. Four different representations are generated for two datasets that contain electric guitar recordings with pedal effects being applied to them, isolated or alongside other instruments, so that the neural network can be trained with the provided datasets being split through the k-fold method to calculate the accuracy of each model. The classifier built using the HCQT managed to perform well and was able to reach accuracy levels of 95.9% in one of the datasets, a value greater than what is achieved by a human expert on this problem, as well as a performance comparable to other recent studies for the same task.
