Speaker: Sonia Aoi García Shida.
Abstract: Automatic composer classification is a challenging task within the field of Music Information Retrieval, as it requires identifying compositional styles between composers from the same musical period, unlike genre classification where differences between classes are more evident. This work proposes a system for classifying a set of classical composers using a subset of the MusicNet dataset, comparing traditional machine learning models against deep learning architectures based on CNNs and CRNNs. The pipeline is first validated on the GTZAN genre classification benchmark before being applied to the main task. We also analyse the contribution of different audio features and evaluate the effect of data augmentation strategies on model performance. Results indicate that both classical models and deep learning architectures achieved similar performance, with a balanced accuracy of around 61.5%, showing that in low-data scenarios classical models can be just as effective as deep learning architectures.