Abstract
Text-to-Audio (TTA) generation has seen significant progress, but evaluating these systems remains challenging. Existing reference-free metrics often focus on global semantic alignment, neglecting fine-grained acoustic events. In this paper, we introduce ELSA (Event-Wise Semantic Alignment), a novel metric designed to evaluate TTA models by aligning acoustic events with their corresponding textual descriptions. Our experiments demonstrate that ELSA correlates better with human judgment compared to state-of-the-art baselines.
Method
The ELSA framework leverages advanced alignment techniques to map specific acoustic events in the generated audio to semantic units in the text prompt. This allows for a more granular evaluation of audio quality and relevance.
Figure 2: The architecture of the ELSA metric.
Interactive Demo
Explore the ELSA pipeline in action. Select a sample to see how the text and audio are processed.
Experimental Results
Swipe to see our comparison with state-of-the-art metrics and ablation studies.
Table 1: Comparison with SOTA metrics.
Table 2: Ablation analysis.
Table 3: Extended evaluation.
Analysis
Detailed analysis of event sensitivity compared to previous approaches.
ELSA Event Sensitivity
ELSA accurately captures fine-grained events.
Previous Approaches
Existing metrics struggle with temporal event alignment.
Citation
To be appeared.