M2CQA-S: Speech-Augmented Multilingual Contrastive Visual Grounding

M2CQA-S is a culturally grounded benchmark for studying whether multimodal models make consistent decisions when equivalent statements are presented as text or speech, in English or Modern Standard Arabic (MSA). Each image is paired with three statements: one visually supported statement and two plausible but unsupported alternatives. Spoken versions include male and female voice variants.

The benchmark accompanies:

Said Aloud, Read Different: Cross-Modal Instability in Multimodal Models
Basel Mousi, Fahim Dalvi, Shammur Chowdhury, Firoj Alam, and Nadir Durrani
Interspeech 2026, pp. 3836–3841.

Paper | arXiv | Code | Related image–text benchmark

Dataset Overview

The benchmark spans 18 countries in the Middle East and North Africa (MENA) and includes culturally distinctive scenes such as architecture, clothing, cuisine, religious settings, and public spaces. It supports comparisons across language, input modality, and voice, while retaining the same image and statement meaning.

Configuration	Language	Split	Rows in the current release
english	English	train	10,151
arabic-msa	Modern Standard Arabic	train	10,151

Each row contains one image, its three statements and labels, and 12 audio references: statement-only and full-prompt recordings for each of the three statements in male and female voices. Both configurations expose the same 27-column schema. The English and Arabic subsets list rows in a different order, so they should be matched on image_id plus country, not by row number if parallelism between samples is required.

Dataset Construction

Image-grounded questions are converted into contrastive statement triplets after image-blind filtering. English statements are translated into Arabic, and speech is synthesized with XTTS-v2 using human reference voices. Human verification covers a stratified 15% subset. See the paper for filtering, annotation, speech-quality checks, and acoustic robustness experiments.

Data Fields

In the table below, {i} stands for 1, 2, or 3.

Field	Type	Description
image	Image	Image used to judge all three statements.
image_id	string	Image identifier.
image_url	string	Source image URL.
country	string	Country associated with the image.
category	string	Broad thematic category.
subcategory	string	Finer thematic category.
cf{i}_statement	string	Statement to be judged against the image.
cf{i}_prompt	string	Full textual task prompt, including instructions and the statement.
cf{i}_is_correct	bool	Ground-truth label: True if the statement is visually supported; False otherwise.
cf{i}_male_statement_audio	Audio	Male-voice recording of the statement.
cf{i}_female_statement_audio	Audio	Female-voice recording of the statement.
cf{i}_male_prompt_audio	Audio	Male-voice recording of the full prompt.
cf{i}_female_prompt_audio	Audio	Female-voice recording of the full prompt.

Do not infer the label from the statement index. Use cf{i}_is_correct: the supported statement is not necessarily cf1.

Usage

Install the Hugging Face dataset libraries, image support, and audio decoding dependencies:

pip install -U "datasets[audio]" huggingface_hub pillow

Load either language configuration:

from datasets import load_dataset

ds = load_dataset("QCRI/M2CQA-S", "english", split="train")
# For Modern Standard Arabic:
# ds = load_dataset("QCRI/M2CQA-S", "arabic-msa", split="train")

print(ds)

Images and audio are stored as external files referenced by hf:// URLs and fetched on access, with local caching. The Parquet tables contain metadata and media references, so downloading the tables alone does not download the media.

Inspect Text and Labels Without Decoding Media

text_columns = ["image_id", "country", "category", "subcategory"]
text_columns += [
    f"cf{i}_{field}"
    for i in (1, 2, 3)
    for field in ("statement", "prompt", "is_correct")
]

example = ds.select_columns(text_columns)[0]
print(example["image_id"], example["country"])

for i in (1, 2, 3):
    print(example[f"cf{i}_statement"])
    print("Supported:", example[f"cf{i}_is_correct"])

Access an Image and Spoken Statement

image = ds.select_columns(["image"])[0]["image"]
audio = ds.select_columns(["cf1_male_statement_audio"])[0][
    "cf1_male_statement_audio"
]

Audio decoding depends on your installed datasets version and its audio backend. Recent versions use TorchCodec and require a compatible FFmpeg installation. To inspect stored references without decoding:

from datasets import Audio

audio_refs = ds.select_columns(["cf1_male_statement_audio"])
audio_refs = audio_refs.cast_column(
    "cf1_male_statement_audio", Audio(decode=False)
)
print(audio_refs[0]["cf1_male_statement_audio"])

For reproducibility, pass a pinned commit hash through revision= when loading or downloading the dataset.

Pre-download for Offline or Bulk Use

One Language Configuration

from huggingface_hub import snapshot_download

config = "english"  # or "arabic-msa"
local_dir = snapshot_download(
    repo_id="QCRI/M2CQA-S",
    repo_type="dataset",
    allow_patterns=[
        "README.md",
        f"data/{config}.parquet",
        "images/**",
        f"audios/{config}/**",
    ],
)

One Country's Media

config = "english"
country = "algeria"

local_dir = snapshot_download(
    repo_id="QCRI/M2CQA-S",
    repo_type="dataset",
    allow_patterns=[
        "README.md",
        f"data/{config}.parquet",
        f"images/{country}/**",
        f"audios/{config}/{country}/**",
    ],
)

This downloads the full metadata table but only the selected country's media. Filter rows by country before accessing media. Stored references retain their Hub URLs; a downloaded snapshot does not automatically rewrite them to local filesystem paths. For fully offline processing with local paths, map the media references to the corresponding files under local_dir.

Full Repository

local_dir = snapshot_download(
    repo_id="QCRI/M2CQA-S",
    repo_type="dataset",
)

Evaluation

Judge each statement as true or false given the image, using its Boolean label. Keep predictions grouped by image and input condition. Report statement-level accuracy alongside triplet accuracy and Contrastive Instability (CI):

CI = 1 - (number of triplets with all three predictions correct)
         / (number of triplets with at least one prediction correct)

Lower CI means less partial inconsistency. Triplets with zero correct predictions are excluded from its denominator. If that denominator is zero, CI is undefined. Report CI with accuracy to retain visibility into complete failures.

License and Intended Use

The repository is released under CC BY-NC-SA 4.0. Provide attribution, use it for noncommercial purposes, and share adaptations under the same license. Source images may carry additional rights or attribution requirements; retain their provenance through image_url.

The benchmark is intended for research on multimodal grounding, contrastive consistency, and multilingual speech–text evaluation. Its synthesized speech and regional coverage should be considered when interpreting results beyond the benchmark.

Citation

If you use M2CQA-S, please cite:

@inproceedings{mousi26_interspeech,
  title     = {{Said Aloud, Read Different: Cross-Modal Instability in Multimodal Models}},
  author    = {Basel Mousi and Fahim Dalvi and Shammur Absar Chowdhury and Firoj Alam and Nadir Durrani},
  year      = {2026},
  booktitle = {{Interspeech 2026}},
  pages     = {3836--3841},
  doi       = {10.21437/Interspeech.2026-1980},
  issn      = {2958-1796},
  url       = {https://www.isca-archive.org/interspeech_2026/mousi26_interspeech.html}
}