Based on the new dataset created and considering a regression approach to the MER task, CNN models are trained separately on audio, lyrics and comments features and using fusions of these models on combinations of audio & lyrics data and audio & comments, to make predictions in the 2D-space of emotion. This concludes with a comparative discussion on these three modalities.

From studying the timeline described by humans when listening to music, emerged the necessity of creating a new dataset, to provide support for recognizing emotions not only from audio features, but also from lyrics and comments features, with annotations extracted from social tags.

The task of recognizing emotion in music, based on MFCCs features extracted from raw audio, was tackled from different perspectives and using Deep Neural Networks, aiming to gain insights into the field of Music Information Retrieval. This project resulted in a collection of methods and results that can further be used as a starting point for a more complex and diverse research.

This project represents the implementation of the Generative Adversarial Network based on multiple dense residual blocks proposed in [Zhang et al., 2020], with slight adaptation of the architecture and training parameters. An experiment and its results are also presented.