Researchers propose a framework that uses audio-aware large language models (ALLMs) as fine-grained judges to verify target event presence and temporal relations in generated audio. This approach addresses the gap where existing text-to-audio models often fail to follow instructions involving multiple sound events and their order.
- The method constructs preference pairs for direct preference optimization using ALLM feedback, which was validated on benchmarks and through human verification.
- A new narrative benchmark called S3Bench is introduced to evaluate multi-event temporal instruction following.
- Experiments demonstrate improvements in event completeness, temporal ordering, and joint instruction-following accuracy across existing benchmarks and S3Bench while maintaining audio quality.
This work provides a mechanism for better instruction-level correctness in text-to-audio generation by leveraging fine-grained feedback rather than relying solely on global similarity metrics.