Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

AImoclips: A Benchmark for Evaluating Emotion Conveyance in Text-to-Music Generation

Created by
  • Haebom

Author

Gyehun Go, Satbyul Han, Ahyeon Choi, Eunjin Choi, Juhan Nam, Jeong Mi Park

Outline

This paper presents AImoclips, a benchmark for evaluating the emotional expressiveness of text-to-music (TTM) systems. Six state-of-the-art TTM systems were used to generate over 1,000 music clips based on 12 emotional intents, and 111 participants were asked to rate each clip's valence and arousal on a 9-point Likert scale. Experimental results showed that commercial systems tended to produce more pleasant music than intended, while open-source systems exhibited the opposite trend. All systems conveyed emotions more accurately when in a high-arousal state, and all systems exhibited a bias toward emotional neutrality. AImoclips provides insight into the emotional expressive characteristics of each model and supports the future development of emotionally congruent TTM systems.

Takeaways, Limitations

•
Takeaways:
◦
Establishing quantitative evaluation criteria for the emotional expression capabilities of TTM systems (AImoclips benchmark).
◦
Comparative analysis of emotional expression characteristics of commercial and open-source TTM systems (commercial systems tend to produce more pleasant emotions, while open-source systems tend to produce emotions contrary to intentions)
◦
Emotional communication is more effective in a state of high arousal.
◦
Confirming the emotional neutrality bias of the TTM system
•
Limitations:
◦
AImoclips benchmarks may be limited to specific emotions and models.
◦
Assessing emotions using only valence and arousal may not fully reflect the diversity of emotions.
◦
Number of participants may be limited (111 people)
◦
Further research is needed to understand the causes and solutions to emotional neutrality bias.
👍
Made with Slashpage