EXPLAINABILITY METHODS IN MUSIC EMOTION RECOGNITION

This thesis explores how explainability can be introduced into Music Emo- tion Recognition (MER) models, which are usually hard to interpret despite their good performance. While many deep learning models can predict the emotional content of music with high accuracy, they often work as black boxes, giving little to no information about how they reach their conclusions. The goal of this work is to make these models more understandable, especially for users who might want to exploit them not just as tools, but also to learn something from them. To do this, the thesis develops and tests two different approaches. The first one is based on musical features—some taken from the literature, and others introduced as a novel contribution. It starts from an existing deep learning framework that uses mid-level features like melodic or rhythmic descriptors to explain predictions, and then expands it by adding simpler, more intuitive features like chords or notes that could be easier to interpret and possibly helpful to composers or researchers. The second approach instead focuses on raw audio data. Here, the idea is to make the model’s internal reasoning perceptible through sound. Using a Vision Transformer (ViT) trained on spectrograms and Layer-wise Relevance Propagation (LRP), this method creates a modified version of the original music where the most relevant parts for the prediction of the task Happy vs. Sad are made louder, allowing the listener to hear which segments influenced the classification most. Even if the two approaches are different, they aim at the same objective: making MER models more transparent and easier to interact with. The hope is that this kind of explainability can help both in research and in creative applications, giving users a better grasp of the models’ internal reasoning processes.