Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AudioGenX: Explainability on Text-to-Audio Generative Models

Hyunju Kang, Geonhee Han, Yoonjae Jeong, Hogun Park

TL;DR

AudioGenX tackles the lack of transparency in text-to-audio generation by introducing a faithful, perturbation-based explainability framework that uses factual and counterfactual reasoning to quantify token-level importance in cross-attention. It defines soft masks over input tokens, optimizes them via a dual-objective loss $ extbf{L} = extbf{L}_{F} + extbf{L}_{CF} + oldsymbol{ ext{alpha}} extbf{L}_{1}( extbf{M}) + oldsymbol{ ext{beta}} extbf{L}_{2}( extbf{M})$, and provides both holistic and interval-focused explanations. Extensive experiments on AudioCaps with PaSST demonstrate that AudioGenX yields faithful explanations (low $Fid_{F}$ and $KL_{F}$) and strong counterfactual sensitivity (high $Fid_{CF}$ and $KL_{CF}$), while delivering compact masks and enabling practical audio editing insights. This work enhances transparency, trust, and debugging capabilities for TAG models, with implications for more controllable and auditable audio generation, though it notes hyperparameter sensitivity and dataset biases as limitations.

Abstract

Text-to-audio generation models (TAG) have achieved significant advances in generating audio conditioned on text descriptions. However, a critical challenge lies in the lack of transparency regarding how each textual input impacts the generated audio. To address this issue, we introduce AudioGenX, an Explainable AI (XAI) method that provides explanations for text-to-audio generation models by highlighting the importance of input tokens. AudioGenX optimizes an Explainer by leveraging factual and counterfactual objective functions to provide faithful explanations at the audio token level. This method offers a detailed and comprehensive understanding of the relationship between text inputs and audio outputs, enhancing both the explainability and trustworthiness of TAG models. Extensive experiments demonstrate the effectiveness of AudioGenX in producing faithful explanations, benchmarked against existing methods using novel evaluation metrics specifically designed for audio generation tasks.

AudioGenX: Explainability on Text-to-Audio Generative Models

TL;DR

AudioGenX tackles the lack of transparency in text-to-audio generation by introducing a faithful, perturbation-based explainability framework that uses factual and counterfactual reasoning to quantify token-level importance in cross-attention. It defines soft masks over input tokens, optimizes them via a dual-objective loss , and provides both holistic and interval-focused explanations. Extensive experiments on AudioCaps with PaSST demonstrate that AudioGenX yields faithful explanations (low and ) and strong counterfactual sensitivity (high and ), while delivering compact masks and enabling practical audio editing insights. This work enhances transparency, trust, and debugging capabilities for TAG models, with implications for more controllable and auditable audio generation, though it notes hyperparameter sensitivity and dataset biases as limitations.

Abstract

Text-to-audio generation models (TAG) have achieved significant advances in generating audio conditioned on text descriptions. However, a critical challenge lies in the lack of transparency regarding how each textual input impacts the generated audio. To address this issue, we introduce AudioGenX, an Explainable AI (XAI) method that provides explanations for text-to-audio generation models by highlighting the importance of input tokens. AudioGenX optimizes an Explainer by leveraging factual and counterfactual objective functions to provide faithful explanations at the audio token level. This method offers a detailed and comprehensive understanding of the relationship between text inputs and audio outputs, enhancing both the explainability and trustworthiness of TAG models. Extensive experiments demonstrate the effectiveness of AudioGenX in producing faithful explanations, benchmarked against existing methods using novel evaluation metrics specifically designed for audio generation tasks.

Paper Structure

This paper contains 24 sections, 15 equations, 9 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. The Proposed AudioGenX
  5. Definition of Masks as Explanations
  6. Formulating Factual Explanations
  7. Formulating Counterfactual Explanations
  8. Objective Function for AudioGenX
  9. Providing Audio-Level Explanations
  10. Experimental Setup
  11. Experimental Results
  12. RQ 1: Does AudioGenX Generate Faithful Explanations?
  13. RQ 2: How Well Do the Explanations from AudioGenX Reflect the Generated Audio?
  14. RQ 3: How Can Explanations Help Understand AudioGen Behavior?
  15. RQ 4: Does AudioGenX Generate Explanations Efficiently?
  16. ...and 9 more sections

Figures (9)

  • Figure 1: A comprehensive explanation provided by AudioGenX for the entire audio in (a). Granular explanations for the interval from 1 to 1.5 seconds in (b) and from 2.5 to 3 seconds in (c), respectively.
  • Figure 2: (a) The process by which AudioGen generates an audio. (b) AudioGenX's procedure for generating and applying explanations, with the $Explainer$ in the green box. (c) The method for calculating and applying the loss in AudioGenX.
  • Figure 3: Visualization of AudioGenX and other methods.
  • Figure 4: Explanation generated by AudioGenX for two audios created from a single prompt. (a) includes bird sounds, while (b) does not.
  • Figure 5: Explanations generated from negated prompts: (a) single negation, (b) double negation.
  • ...and 4 more figures