---
pretty_name: RMS-AQA
task_categories:
  - question-answering
language:
  - en
tags:
  - audio-question-answering
  - spatial-audio
  - first-order-ambisonics
  - benchmark
  - icassp-grand-challenge
extra_gated_prompt: Please login HuggingFace to register your email and research affiliation to get auto-approval.
extra_gated_fields:
  Full Name (required):
    type: text
  Email (required):
    type: text
  Affiliation (required):
    type: text
  Country:
    type: country
  I want to use this dataset for:
    type: select
    options:
      - Research
      - Education
      - Product Development
      - label: Other
        value: other
---

# RMS-AQA: Real-World Multi-Hop Spatial Audio Question Answering Challenge

[![Code Repository](https://img.shields.io/badge/RMS--AQA-Code%20Repository-444444?logo=github&logoColor=white)](https://github.com/rmsaqachallenge/rmsaqa-code)
[![Challenge Website](https://img.shields.io/badge/RMS--AQA-Challenge%20Website-0b7285?logo=githubpages&logoColor=white)](https://rmsaqachallenge.github.io)

**A spatial audio question answering (SAQA) dataset for real-world domestic environments.**

## Overview

**RMS-AQA** is a spatial audio question answering benchmark built from 10-second First-Order Ambisonics (FOA) recordings collected in smart-home environments. The benchmark is organized as a two-stage question-answering task:

1. **Stage 1:** ground the audible sound events present in the scene.
2. **Stage 2:** perform complex spatio-temporal reasoning based on the Stage 1 answer as context.

All audio clips are provided in four-channel, **24 kHz, 10-second FOA format**, and every question answering (QA) pair is a multiple-choice question **(MCQ)** generated by **GPT-5.5** and paraphrased by **Gemini 3.1**. The dataset is split into training, validation, and test sets: the training set is fully synthetic, the validation set combines synthetic and recorded examples, and the hidden evaluation set consists entirely of recorded data. Simulated audio is generated with **SpatialScaper** by convolving dry sound events with measured spatial room impulse responses and the recorded audios are captured with a **Zoom H3-VR** recorder. 

Stage 2 covers six question dimensions:

- **Sound counting (SC)**
- **Spatial location (SL)**
- **Temporal detection (TD)**
- **Temporal relation (TR)**
- **Spatial relation (SR)**
- **Action prediction (AP)**

## Dataset Structure

Each record contains a 10-second audio segment and its two-stage MCQ pair. Stage 1 identifies the sound events present in the scene; Stage 2 asks a complex spatio-temporal question conditioned on the Stage 1 context.The train set contains **170K** clips and **170K** QA pairs and the validation set is made up of **5K** audio clips with **6K** QA pairs. The hidden test set includes **2K** clips with **3K** QA pairs, which will be released on **Nov. 24, 2026**.

### QA example

```json
{
  "records": [
    {
      "segment_id": "train-000000.json",
      "qa_pairs": [
        {
          "stage": 1,
          "question": "What types of sound events can be heard in this audio?",
          "answer": "D",
          "options": {
            "A": "The recording includes key drop.",
            "B": "In the audio, you can hear dishwasher.",
            "C": "The clip contains indoor cricket chorus.",
            "D": "The audible sound types are child shouting."
          }
        },
        {
          "stage": 2,
          "question": "If a domestic robot were monitoring this audio, what should its response plan be?",
          "answer": "D",
          "options": {
            "A": "The best next step is to log and continue passive monitoring for dishwasher near the front-left area.",
            "B": "For this clip, the embodied assistant should trigger a home emergency alert, notify, and monitor safely for child shouting near the front-left area.",
            "C": "The appropriate response plan is to trigger a home emergency alert, notify, and monitor safely for dishwasher near the rear area.",
            "D": "Because of the detected sound, the robot should make a non-intrusive check and notify for child shouting near the rear area."
          },
          "category_id": 6,
          "category_name": "Action Prediction"
        }
      ]
    }
  ]
}
```

## Download

With [huggingface_hub](https://huggingface.co/docs/huggingface_hub) (recommended):

```bash
pip install huggingface_hub
huggingface-cli download PeacefulData/RMS-AQA --repo-type dataset --local-dir ./RMS-AQA
```

With Python:

```python
from huggingface_hub import snapshot_download

snapshot_download(repo_id="PeacefulData/RMS-AQA", repo_type="dataset", local_dir="./RMS-AQA")
```

Or with git (requires [git-lfs](https://git-lfs.com)):

```bash
git clone https://huggingface.co/datasets/PeacefulData/RMS-AQA
```

## Folder Layout

```text
RMS-AQA/
├── train_QA/                  # training set QA pairs (train_QA.tar.zst)
├── train_audio/               # training set audio shards (train-*.tar.zst)
├── dev_QA/                    # development set QA pairs (dev_QA.tar.zst)
└── dev_audio/                 # development set audio shards (dev-*.tar.zst)
```

The `*_QA` archives contain per-sample JSON files in the schema shown above; the `*_audio` archives contain the corresponding four-channel FOA clips.