Estimativa de andamento musical através de escalogramas wavelet e redes neurais convolucionais

Abstract

Audio tempo estimation is one of the most fundamental tasks in Music Information Retrieval (MIR). In this work, a wavelet scalogram is used as a two-dimensional image representation of the audio signal. Different ways of generating the wavelet scalogram were tested by varying the mother wavelet function and scale levels. The images were used to train a Convolutional Neural Network (CNN) through supervised learning, relating the image to a target tempo value. The k-fold cross-validation method was used to produce greater statistical reliability of the proposed model and to define the best result for choices involving the parameters of scalogram generation. Data augmentation was implemented online, modifying the scalograms during training. Finally, the model was evaluated on widely used databases in the literature, and the results were compared to the state-of-the-art. Results compatible with state-of-the-art were achieved on the "GiantSteps" evaluation database achieving an accuracy (Type 2 - ACC2) of 92.6\% with the Morlet and Shannon mother wavelets.