Automatic Emotion and Theme Recognition with a Convolutional Neural Network
Emotion and theme recognition are common problems in music information retrieval. The listeners' current emotional state influences musical choice significantly, therefore, emotion recognition could improve automatic song recommendations. The 2019 Emotion and Theme Recognition in Music Task by the MediaEval Benchmarking Initiative is a yearly challenge designed to contribute to solving automatic emotion and theme recognition. This bachelor project aims to improve upon some solutions which were handed in for 2019. We try to achieve this by utilizing a convolutional neural network and feeding mel-spectograms of the songs into it. The data for training and evaluation is extracted from the Acousticbrainz dataset. We utilize the low-level data, which had been generated using the Essentia toolkit. The results are evaluated using ROC-AUC, PR-AUC, F1-score (macro), and F1-score (micro) for better comparability with the previous task solutions.