AudioGenX: Explainability on Text-to-Audio Generative Models

Text-to-audio models lack transparency: Current AI systems that generate audio from text descriptions don't clearly show how specific words in the description influence the final audio output, creating a "black box" problem that limits user trust and understanding.

AudioGenX identifies which words matter most: The method highlights which input words are most important for generating particular sounds, working at the detailed level of individual audio tokens to provide granular explanations of the generation process.

Faithful explanations require dual objectives: The approach uses both "factual" explanations (what the model actually uses) and "counterfactual" explanations (what would happen if certain words were removed) to ensure the explanations genuinely reflect how the model works, rather than just appearing plausible.

Explainability improves trustworthiness: By making text-to-audio generation models more transparent and understandable, AudioGenX helps users and developers better trust these systems and identify potential biases or errors in how audio is being generated from text descriptions.

New evaluation methods for audio AI: The research introduces novel metrics specifically designed to test whether explanations for audio generation are actually accurate, addressing a gap in how we currently measure explainability in this emerging technology.