Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Audio Hallucinations in Large Audio-Video Language Models

Taichi Nishimura, Shota Nakada, Masayoshi Kondo

TL;DR

This work investigates audio hallucinations in large audio–video language models, focusing on Video LLAMA's tendency to describe visual content while ignoring audio cues. It builds a labeled corpus from 1,000 prompts, annotating hallucination presence and three types, and finds about $32.3\%$ of sentences hallucinated, with type (C) most frequent. The authors evaluate audio-hallucination detection using pre-trained audio–text embeddings in zero-shot and fine-tuning settings, using MS-CLAP and LAION-CLAP as backbones; zero-shot yields about $52.2\%$ F1, while fine-tuning reaches $87.9\%$ F1, with MS-CLAP generally outperforming LAION-CLAP. The study highlights the value of explicit audio–text alignment to improve faithful audio grounding in multimodal LLMs and offers practical classifiers and insights for reducing auditory hallucinations in video descriptions.

Abstract

Large audio-video language models can generate descriptions for both video and audio. However, they sometimes ignore audio content, producing audio descriptions solely reliant on visual information. This paper refers to this as audio hallucinations and analyzes them in large audio-video language models. We gather 1,000 sentences by inquiring about audio information and annotate them whether they contain hallucinations. If a sentence is hallucinated, we also categorize the type of hallucination. The results reveal that 332 sentences are hallucinated with distinct trends observed in nouns and verbs for each hallucination type. Based on this, we tackle a task of audio hallucination classification using pre-trained audio-text models in the zero-shot and fine-tuning settings. Our experimental results reveal that the zero-shot models achieve higher performance (52.2% in F1) than the random (40.3%) and the fine-tuning models achieve 87.9%, outperforming the zero-shot models.

On the Audio Hallucinations in Large Audio-Video Language Models

TL;DR

This work investigates audio hallucinations in large audio–video language models, focusing on Video LLAMA's tendency to describe visual content while ignoring audio cues. It builds a labeled corpus from 1,000 prompts, annotating hallucination presence and three types, and finds about of sentences hallucinated, with type (C) most frequent. The authors evaluate audio-hallucination detection using pre-trained audio–text embeddings in zero-shot and fine-tuning settings, using MS-CLAP and LAION-CLAP as backbones; zero-shot yields about F1, while fine-tuning reaches F1, with MS-CLAP generally outperforming LAION-CLAP. The study highlights the value of explicit audio–text alignment to improve faithful audio grounding in multimodal LLMs and offers practical classifiers and insights for reducing auditory hallucinations in video descriptions.

Abstract

Large audio-video language models can generate descriptions for both video and audio. However, they sometimes ignore audio content, producing audio descriptions solely reliant on visual information. This paper refers to this as audio hallucinations and analyzes them in large audio-video language models. We gather 1,000 sentences by inquiring about audio information and annotate them whether they contain hallucinations. If a sentence is hallucinated, we also categorize the type of hallucination. The results reveal that 332 sentences are hallucinated with distinct trends observed in nouns and verbs for each hallucination type. Based on this, we tackle a task of audio hallucination classification using pre-trained audio-text models in the zero-shot and fine-tuning settings. Our experimental results reveal that the zero-shot models achieve higher performance (52.2% in F1) than the random (40.3%) and the fine-tuning models achieve 87.9%, outperforming the zero-shot models.
Paper Structure (16 sections, 1 equation, 7 figures, 2 tables)

This paper contains 16 sections, 1 equation, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Audio captioning
  4. Hallucinations in LLMs
  5. Audio hallucination analysis
  6. Annotation
  7. Statistics on the constructed dataset
  8. Qualitative evaluation
  9. Audio hallucination classification
  10. Zero-shot approach
  11. Fine-tuning approach
  12. Experiments
  13. Experimental settings
  14. Quantitative evaluation
  15. Qualitative evaluation
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Examples of audio hallucinations. We collect 1,000 response sentences from Video LLAMA and annotate whether the sentences are hallucinated or not. If a sentence is hallucinated, we annotate the hallucination types with it.
  • Figure 2: Statistics on the hallucination types.
  • Figure 3: Audio hallucination examples for each type.
  • Figure 4: Audio hallucination cases on musical instruments.
  • Figure 5: Zero-shot audio hallucination classifier.
  • ...and 2 more figures