README.md

---
language:
- en
- zh
- ru
- th
- bn
license: cc-by-nc-4.0
multilinguality:
- multilingual
pretty_name: GlobeAudio
task_categories:
- audio-text-to-text
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 44100
  - name: question
    dtype: string
  - name: option1
    dtype: string
  - name: option2
    dtype: string
  - name: option3
    dtype: string
  - name: option4
    dtype: string
  - name: target
    dtype: string
  download_size: 1554418131
  dataset_size: 1554462926
tags:
- audio
- in-the-wild
- compound-task
- natural
- naturalistic
- multilingual
- multicultural
configs:
- config_name: en
  data_files:
  - split: train
    path: data/en/en-*
  default: true
- config_name: zh
  data_files:
  - split: train
    path: data/zh/zh-*
- config_name: sg
  data_files:
  - split: train
    path: data/sg/sg-*
- config_name: ru
  data_files:
  - split: train
    path: data/ru/ru-*
- config_name: th
  data_files:
  - split: train
    path: data/th/th-*
- config_name: bn
  data_files:
  - split: train
    path: data/bn/bn-*
size_categories:
- 1K<n<10K
---

# GlobeAudio Dataset

This is the dataset for the paper [GlobeAudio: A Multilingual Multicultural Benchmark for Naturalistic Evaluation of Large Audio-Language Models](https://huggingface.co/papers/2606.08194).

## Table of Contents
- [Dataset Description](#dataset-description)
- [Language Selection](#language-selection)
- [Dataset Quality](#dataset-quality)
- [Dataset Breakdown](#dataset-breakdown)
- [Sample Usage](#sample-usage)


## Dataset Description
GlobeAudio is a multilingual, multicultural benchmark for assessing naturalistic audio understanding, comprising 5,637 human-authored and rigorously verified MCQs across six typologically diverse languages, namely English (United States), Chinese (China), Thai (Thailand), Russian (Russia), Bengali (India) and Singlish (Singapore). The dataset comprises of naturally occurring audio clips from online media that reflect real-life speech across diverse contexts, and are curated to rigorously examine the ability of multilingual systems to handle acoustic and sociolinguistic variability.

## Language Selection
The 6 carefully selected languages span multiple language families (Indo-European, Sino-Tibetan, Tai-Kadai and creole varieties), covers a wide range of geographic regions, and reflects substantial variation in resource availability and phonological characteristics. They also differ markedly in script systems and spoken characteristics -- the inclusion of Singlish enables the evaluation of naturally occuring code-mixing and multilingual speech within a single variety, capturing linguistic phenomena that are common in real-world audio but rarely represented in existing benchmarks.

## Dataset Quality
For all selected languages, data annotation is performed exclusively by native speakers for whom the language is their first language, in order to ensure high linguistic fidelity, culturally appropriate interpretation and reliable construction of questions and distractors. Consequently, a two-stage quality control process was conducted involving random sampling and cross-checking, producing a final dataset with an inter-annotator agreement of 95.5%, together with verified and consistent gold labels.

## Dataset Composition

| Language (Code) | CC Size | Isochrony | Genus      | Script   | Clip Length (s) | QA Length (Q/A) | Total |
|-----------------|---------|-----------|------------|----------|-----------------|-----------------|-------|
| English (en)    | 42.60   | High, Str | Germanic   | Latin    | 25.72           | 53.8 / 22.9     | 1,274 |
| Russian (ru)    | 6.15    | High, Str | Slavic     | Cyrillic | 25.69           | 37.8 / 22.2     | 924   |
| Chinese (zh)    | 4.99    | Mid, Syl  | Sinitic    | Hanzi    | 24.88           | 13.5 / 5.0      | 1,072 |
| Thai (th)       | 0.37    | Mid, Syl  | Tai        | Thai     | 24.82           | 29.3 / 13.0     | 1,145 |
| Bengali (bn)    | 0.10    | High, Syl | Indo-Aryan | Bengali  | 25.47           | 37.6 / 18.3     | 719   |
| Singlish (sg)   | --      | Low, Syl  | --         | Latin    | 25.54           | 73.3 / 34.9     | 503   |
| **Total**       |         |           |            |          | **25.35**       | **40.9 / 19.4** | **5,637** |

> *Isochrony is categorized by syllable complexity and rhythmic unit (Str = stress-timed, Syl = syllable-timed).*

Each audio clip is approximately 20-40 seconds long, and contains naturally occurring audio that reflect real-life auditory settings spanning both formal and informal contexts.

## Sample Usage
The `datasets` library lets you load and preprocess datasets in Python at scale. You may load the dataset to your local drive with `load_dataset`.

For example, to download the English config, specify the corresponding language config name (i.e., "en" for English):
```python
from datasets import load_dataset

# load a specific language eg. "en", "ru", "zh", "sg", "th", "bn"
dataset = load_dataset("iNLP-Lab/GlobeAudio", "en")
```