Advancing guitar emotion recognition through audio data augmentation to enhance smart musical instruments

Dataset augmentation techniques have been widely used to achieve state-of-the-art results in Music Information Retrieval tasks. However, their application in music emotion recognition (MER) remains underexplored. MER methods are particularly relevant to the design of smart musical instruments (SMIs), as emotionally aware SMIs have the potential to enrich musical interaction by providing feedback to musicians or dynamically adjusting their sound properties. In this study, we analyze the effect of 11 augmentation techniques on emotion classification in guitar recordings using a convolutional neural network. Our dataset consists of approximately 400 guitar recordings labeled with four emotions: aggressiveness, relaxation, happiness, and sadness. Results indicate that time shift, time stretch, and pitch shift provide the most significant improvements in classification accuracy. Further analysis combining these techniques under different settings yielded similar performance outcomes. A listening test confirmed that the applied augmentations did not significantly alter the perceived emotional content of the recordings. These findings support the development of emotionally aware SMIs by enhancing MER accuracy through data augmentation, ultimately enabling more expressive and interactive music-making experiences.