ELSA: Acoustic Event-Level Semantic Alignment for Fine-Grained Reference-Free Text-to-Audio Evaluation

2026-06-15

AI Summary

Existing automatic evaluation metrics for text-to-audio generation exhibit weak correlation with human judgments in fine-grained semantic alignment. To address this limitation, this work proposes ELSA—an event-level, reference-free evaluation method that introduces an acoustic event decomposition mechanism to extract key sound events from the input text and assess semantic alignment between the generated audio and the text at the event level, thereby overcoming the coarse-grained constraints of conventional holistic similarity measures. Built upon the CLAP framework, ELSA demonstrates substantially stronger correlation with human subjective ratings than existing automatic metrics across four text-to-audio generation benchmarks.

Abstract

Text-to-audio (TTA) generation, synthesizing audio from natural language, has been widely studied for its ability to capture precise user intent. To effectively advance TTA models, it is essential to reliably evaluate generated audio without relying on costly human subjective ratings, motivating the development of automatic evaluation metrics that correlate well with human judgments. While recent CLAP-based metrics provide practical reference-free solutions, their coarse-grained text-audio similarity matching often correlates poorly with human ratings. To address this, we propose ELSA, a reference-free evaluation metric for fine-grained text-audio alignment. ELSA decomposes generated audio guided by distinct acoustic events derived from the text query and assesses event-level alignment. Experiments across four TTA benchmarks show that ELSA reveals a higher correlation with human subjective ratings than prior metrics, highlighting its effectiveness for reliable TTA evaluation.

Problem

• text-to-audio evaluation
• reference-free metric
• acoustic event alignment
• human correlation
• fine-grained alignment

Innovation

• event-level alignment
• fine-grained evaluation
• reference-free metric
• text-to-audio generation
• acoustic event decomposition