Τεχνικές βαθιάς μηχανικής μάθησης για αναγνώριση μουσικού συναισθήματος
View/
Keywords
Βαθιά μηχανική μάθηση ; Μεταφορά μάθησης ; Επαύξηση δεδομένων ; StyleGAN2-ADA ; Συνελικτικά νευρωνικά δίκτυα ; Ταξινόμηση συναισθήματος ; DL ; CNN ; Transfer learning ; Music emotion recognition ; AIAbstract
Music is a carrier of many and strong emotions. With the development of technology and the Internet, access to a huge amount of music content is instantaneous from almost anywhere. Despite the availability, choosing music based on the listener's emotion state is quite a difficult task.
This work investigates the ability of well-known CNN architectures (VGG, AlexNet, DenseNet, Inception, ResNeXt, SqueezeNet) to recognize musical emotion when data is limited, with training sets of different distributions and not always balanced. We used Deep Learning techniques such as Transfer Learning and Data Augmentation with Generative Adversarial Networks (GANs).
Using “classic” machine learning, handcrafted features of all audio samples were extracted and classifiers were trained (SVM, K-NN, Random Forest, Extra Trees) in order to have a reference point for the results.
Then, the samples were converted to Mel-spectrograms as inputs to CNNs which were trained running two Transfer Learning scenarios and gave models that were tested in emotion classification experiments. Finally, using StyleGAN2-ADA, we did data augmentation and a new artificial set was created, and also tested in classification tasks.
As ground truth for the experiments, we used the 360-set of Eerola & Vuoskoski's research fully labeled by music experts, which makes it a quite rare set. It consists of 360 soundtrack excerpts from 15'' to 30'' duration, classified into Energy (high, medium, low), Valence (positive, neutral, negative), Tension (high, medium, low) and Emotions (anger, fear, happy, sad, tender). As far as we know, this is the first work that carries out such extensive experiments in this set.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
Ινστιτούτο Πληροφορικής και Τηλεπικοινωνιών του ΕΚΕΦΕ «Δημόκριτος»Number of pages
133Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα