Semi-supervised emotion-driven music generation model based on category-dispersed Gaussian Mixture Variational...

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Machine Learning

Background:

• Current emotion-driven music generation models often require extensive labeled data, limiting their practical application.
• Existing methods lack transparency and precise control over emotional expression in generated music.
• Interpretability and controllability are key challenges in developing advanced music generation systems.

Purpose of the Study:

• To propose a semi-supervised emotion-driven music generation model that overcomes limitations of existing approaches.
• To enhance the interpretability and controllability of emotions in generated music.
• To improve the association between music and emotions through latent space manipulation.

Main Methods:

• Development of a controllable music generation model that disentangles and manipulates rhythm and tonal features.
• Implementation of a semi-supervised model using category-dispersed Gaussian mixture variational autoencoders (GMM-VAEs) for emotion inference.
• Optimization of an objective loss function to improve the separation of emotional clusters in the latent space.

Main Results:

• The proposed model effectively separates music with different emotions within the latent space.
• Experimental results confirm a strengthened association between music and emotions.
• Successful disentanglement and separation of various musical features, enabling accurate emotion-driven generation and transitions.

Conclusions:

• The semi-supervised GMM-VAE approach offers a robust solution for controllable and interpretable emotion-driven music generation.
• The model's ability to disentangle features facilitates nuanced control over musical emotion and transitions.
• This research advances the field of affective computing in music generation by improving data efficiency and model control.