SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models

TL;DR: Speech content is not necessarily visual evidence. SVHalluc is the first benchmark that tests whether audio-visual large language models (AV-LLMs) can distinguish what is said from what is actually seen.

SVHalluc was introduced in “SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models,” CVPR 2026.

• Project page
• Paper (arXiv)
• Code

Contact

For any questions about the dataset or code, please contact Chenshuang Zhang (chenshuang.zhang02@gmail.com).

Motivation

Existing audio-visual hallucination benchmarks commonly treat environmental sounds, such as a dog barking, as indicators that an event occurred. Human speech is fundamentally different: it contains dense semantic information and can refer to past, present, or future events. An AV-LLM may therefore recognize the speech correctly while still incorrectly assuming that the spoken content is visible.

SVHalluc evaluates this distinction from two complementary perspectives:

• Semantic hallucination: What content mentioned in speech is supported by the video?
• Temporal hallucination: When does the narrated event happen visually?

Benchmark Overview

The released test set contains 2,405 video-question pairs:

Category	Task	Abbreviation	Description
Semantic	Global Semantic Alignment	GSA	Video-level semantic matching between speech and visual events
Semantic	Fine-grained Semantic Alignment	FGSA	Object-level semantic grounding
Semantic	Cross-Modal Semantic Binding	CSMB	Event-level semantic grounding
Temporal	Temporal Alignment	TA	Whether the narrated event and visual event occur at the same time
Temporal	Temporal Forecasting	TF	Whether the narrated event occurs in the past, present, or future relative to the speech
Temporal	Cross-Modal Temporal Binding	CMTB	Which visual action occurs when the speech is heard

Task Definitions

Semantic tasks

1. Global Semantic Alignment (GSA) asks whether the speech describes the visual events at the video level.
2. Fine-grained Semantic Alignment (FGSA) asks whether a specific object mentioned by speech is visually present.
3. Cross-Modal Semantic Binding (CSMB) asks whether a specific narrated event is visible.

Temporal tasks

1. Temporal Alignment (TA) asks whether the narrated event is happening at the same time that its speech is heard.
2. Temporal Forecasting (TF) asks whether the narrated event occurs in the past, present, or future relative to the speech.
3. Cross-Modal Temporal Binding (CMTB) asks the model to select the visual action occurring when the speech is heard.

Data Format

The repository has the following structure:

svhalluc/
├── README.md
├── samples.json
├── semantic/
│   └── .../*.mp4
└── temporal/
    └── .../*.mp4

samples.json is a JSON list. Each element represents one video-question pair:

{
  "video_path": "semantic/119/wqpqx-Qm7lk/e4714d74c73bbe3b.mp4",
  "data_type": "semantic",
  "task_type": "Global Semantic Alignment",
  "task_type_abbr": "GSA",
  "question": "Does the speech describe the visual events in the video? (A) Yes (B) No",
  "answer": "B"
}

Fields

Field	Type	Description
video_path	string	Path to the MP4 file, relative to the repository root
data_type	string	High-level category: semantic or temporal
task_type	string	Full task name
task_type_abbr	string	Task abbreviation: GSA, FGSA, CSMB, TA, TF, or CMTB
question	string	Multiple-choice question, including its answer options
answer	string	Correct option letter (A, B, or C)

Multiple questions may reference the same video. Consumers should therefore use video_path rather than row index as the video identifier.

Loading the Dataset

With Python

import json
from pathlib import Path

root = Path("/path/to/svhalluc")
samples = json.loads((root / "samples.json").read_text(encoding="utf-8"))

sample = samples[0]
video_path = root / sample["video_path"]
question = sample["question"]
answer = sample["answer"]

Downloading from the Hugging Face Hub

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="zcs15/SVHalluc",
    repo_type="dataset",
)

Evaluation

SVHalluc is designed as a multiple-choice benchmark. A model receives the video (including its audio) and the question exactly as provided in samples.json. The predicted option letter is compared with answer.

Intended Use

SVHalluc is intended for:

• evaluating speech-grounded video understanding;
• diagnosing semantic and temporal hallucination in AV-LLMs;
• comparing multimodal grounding and reasoning methods;

License and Usage Terms

The dataset is created based on YouCook2 dataset which include Youtube videos. Users are responsible for reviewing and complying with the applicable SVHalluc, YouCook2, Youtube terms before downloading, redistributing, or using the data.

The dataset is provided for research-only purposes. No ownership of third-party video content is transferred by this release. If you are a rights holder and believe content should be removed, please contact the dataset author by email: chenshuang.zhang02@gmail.com.

Citation

If you find this dataset useful, please cite:

@InProceedings{Zhang_2026_SVHalluc,
  author    = {Chenshuang Zhang and Kyeong Seon Kim and Chengxin Liu and Tae-Hyun Oh},
  title     = {SVHalluc: Benchmarking Speech--Vision Hallucination in Audio-Visual Large Language Models},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026}
}