Listening without Looking: Modality Bias in Audio-Visual Captioning

TL;DR

The paper addresses whether audio-visual captioning models truly balance information from both modalities or rely predominantly on one. It introduces a rigorous modality robustness protocol and demonstrates a strong audio bias in LAVCap when trained on AudioCaps. To counter this, it proposes AudioVisualCaps, a semi-automatic dataset that provides joint audio-visual captions, and shows that training on this dataset yields more balanced multimodal usage and robustness. The work offers a practical path to fairer evaluation and improved multimodal integration in captioning systems, with potential benefits for downstream multimodal understanding tasks.

Abstract

Audio-visual captioning aims to generate holistic scene descriptions by jointly modeling sound and vision. While recent methods have improved performance through sophisticated modality fusion, it remains unclear to what extent the two modalities are complementary in current audio-visual captioning models and how robust these models are when one modality is degraded. We address these questions by conducting systematic modality robustness tests on LAVCap, a state-of-the-art audio-visual captioning model, in which we selectively suppress or corrupt the audio or visual streams to quantify sensitivity and complementarity. The analysis reveals a pronounced bias toward the audio stream in LAVCap. To evaluate how balanced audio-visual captioning models are in their use of both modalities, we augment AudioCaps with textual annotations that jointly describe the audio and visual streams, yielding the AudioVisualCaps dataset. In our experiments, we report LAVCap baseline results on AudioVisualCaps. We also evaluate the model under modality robustness tests on AudioVisualCaps and the results indicate that LAVCap trained on AudioVisualCaps exhibits less modality bias than when trained on AudioCaps.

Figures (3)

• Figure 1: Caption examples under modality robustness tests.
• Figure 2: Annotation pipeline for the test set. We first generate visual captions from video frames. Next, an LLM integrates them with audio-focused captions from AudioCaps. Finally, we manually verify the results to produce the final captions that capture both visual and audio modalities.
• Figure 3: Samples from the AudioVisualCaps test set. In contrast to AudioCaps, our captions capture both audio and visual information (visual content is highlighted in green).