ACL2025
T2A-Feedback: Improving Basic Capabilities of Text-to-Audio Generation via Fine-grained AI Feedback
Zehan Wang, Ke Lei, Chen Zhu, Jiawei Huang, Sashuai Zhou, Luping Liu, Xize Cheng, Shengpeng Ji, Zhenhui Ye, Tao Jin, Zhou Zhao
摘要
Figure 1 : The audio description is from a classic Chinese essay "Kou Ji", which vividly depicts a performer using only vocal mimicry to recreate an entire dramatic scene. The existing Text-to-Audio generation model struggles to generate such narrative and multi-event audios. The generated audio often fails to contain all events in the described sequence while maintaining acoustic quality and harmony.