SO-Dataset: Spatial FOA Audio Dataset

SO-Dataset is a large-scale spatial audio dataset in first-order ambisonics (FOA) format. Each example contains FOA waveform and spatial event annotations in DCASE-style CSV files. The dataset combines simulated spatial scenes and real FOA recordings, and all sound event labels are mapped into a unified 63-class sound event taxonomy based on the FSD50k dataset.

The public release stores audio and annotations as tar shards. The tar files preserve the same relative paths used by the metadata, so extracting the archives recreates the audio/ and annotations/ directories expected by the JSONL files.

Dataset Contents

• Audio format: FOA waveform files (.wav)
• Spatial annotations: DCASE-style CSV files
• Labels: unified FSD50k label set
• Splits: train, valid, test
• Metadata: one JSON object per audio scene
• Packaging: path-preserving tar shards for easier download and upload

File Structure

SO-Dataset/
  so_vocab.csv
  tar_shard_summary.json

  metadata/
    train.jsonl
    valid.jsonl
    test.jsonl

  manifests/
    audio-train.jsonl
    audio-valid.jsonl
    audio-test.jsonl
    annotations-train.jsonl
    annotations-valid.jsonl
    annotations-test.jsonl

  archives/
    audio/
      train/
        audio-train-000000.tar
        audio-train-000001.tar
        ...
      valid/
      test/
    annotations/
      train/
        annotations-train-000000.tar
      valid/
        annotations-valid-000000.tar
      test/
        annotations-test-000000.tar

After extraction, the archives create:

audio/{train,valid,test}/*.wav
annotations/{train,valid,test}/*.csv

These paths match the paths stored in metadata/*.jsonl.

Dataset Statistics

SO-Dataset contains 400K FOA audio segments across 233 scenes, with a total of 1.27M annotated sound events.

The dataset and annotations details are shown in the following figure. Figure(a) shows the sub-tasks in SO-QA and SO-Bench, including Detection and Localization, Spatial Relation Understanding, and Complex Reasoning with Semantics. Figure(b) shows the data source of sound events in the dataset. Figure(c) shows the building process of the dataset, including the recording, simulation and collect subset. After building the SO-Dataset, we generate QA pairs and build SO-QA using the metadata of SO-Dataset. Figure(d) shows the distribution of spatial event in our dataset, including azimuth, elevation and distance.

Metadata Format

Each line in metadata/{split}.jsonl is a JSON object describing one FOA scene.

Example:

{
  "schema_version": "spatial_foa_scene_v1",
  "split": "train",
  "dataset": "sim_static",
  "data_source": "sim_static",
  "scene_id": "train/ov2_000000",
  "audio": {
    "duration_seconds": 20.0,
    "foa_path": "audio/train/foa_fed1992f629ae5f3db28.wav"
  },
  "scene_annotation_csv_path": "annotations/train/foa_fed1992f629ae5f3db28.csv",
  "sources": [
    {
      "source_id": "1",
      "track_id": 0,
      "original_label": "telephone_alarm",
      "label": "telephone_alarm",
      "label_id": 50,
      "active_duration_seconds": 19.999937,
      "active_times": [[0.000063, 20.0]],
      "source_trajectory_csv_path": "annotations/train/foa_fed1992f629ae5f3db28_src00.csv",
      "motion": {
        "is_moving": false,
        "pattern": null,
        "description": null,
        "scene_description": null
      }
    }
  ]
}

Important fields:

• audio.foa_path: relative path to the FOA waveform after extraction.
• audio.duration_seconds: audio duration in seconds.
• scene_annotation_csv_path: combined scene-level DCASE-style CSV annotation.
• sources: list of individual sound sources in the scene.
• sources[*].source_trajectory_csv_path: per-source trajectory CSV.
• sources[*].original_label: label before mapping.
• sources[*].label and sources[*].label_id: final FSD63 label name and integer class id. The integer matches so_vocab.csv row order (frequency-sort, 0..62). The loader joins by label string; the integer is kept consistent for users with custom pipelines.
• sources[*].active_times: one or more [start_seconds, end_seconds] intervals.
• sources[*].track_id: source/event track id used in the CSV annotation.

Source-class Vocabulary

so_vocab.csv defines the 63-class sound event taxonomy:

label_id,final_label,count
0,wind_instrument,3841
1,string_instrument,2921
2,guitar,2105
...
62,frog,74

• Row order = SO-Encoder cls-head dimension. The pretrained SO-Encoder checkpoint's classification head was trained with this exact ordering (FSD50K frequency-descending). Do not re-sort the rows.
• The label string in each metadata/*.jsonl source is the authoritative annotation. The dataset loader joins records to vocabulary rows by label name (string), so the integer label_id is informational — but it is also kept consistent with so_vocab.csv (frequency-sort, 0..62) so that users wiring their own pipelines see one and only one class-id system.
• count is the per-class total event count from FSD50K, retained so the ordering is auditable (rows appear in monotonically descending count).

The figure below shows the dataset's per-class event distribution.

Download

Install the Hugging Face CLI:

pip install -U "huggingface_hub[cli]"

Download the full dataset:

hf download dieKarotte/SO-Dataset \
  --repo-type dataset \
  --local-dir SO-Dataset

Download only metadata, manifests, and the label mapping:

hf download dieKarotte/SO-Dataset \
  --repo-type dataset \
  --local-dir SO-Dataset \
  --include "metadata/*" \
  --include "manifests/*" \
  --include "so_vocab.csv" \
  --include "tar_shard_summary.json"

Download only the training audio shards:

hf download dieKarotte/SO-Dataset \
  --repo-type dataset \
  --local-dir SO-Dataset \
  --include "archives/audio/train/*" \
  --include "metadata/train.jsonl" \
  --include "so_vocab.csv"

Extraction

Extract all audio and annotation shards from the dataset root:

cd SO-Dataset

find archives/audio -name "*.tar" -print0 | xargs -0 -n 1 -P 4 tar -xf
find archives/annotations -name "*.tar" -print0 | xargs -0 -n 1 -P 4 tar -xf

Extract only the training split:

cd SO-Dataset

find archives/audio/train -name "*.tar" -print0 | xargs -0 -n 1 -P 4 tar -xf
find archives/annotations/train -name "*.tar" -print0 | xargs -0 -n 1 -P 4 tar -xf

After extraction, metadata paths are directly usable:

audio/train/foa_....wav
annotations/train/foa_....csv
annotations/train/foa_...._src00.csv

Manifests

The files in manifests/ list the tar shards for each group and split.

Example row:

{
  "kind": "audio",
  "split": "train",
  "shard": "archives/audio/train/audio-train-000000.tar",
  "files": 1741,
  "payload_bytes": 4999665664,
  "estimated_tar_bytes": 5000501760
}

Video

As mentioned in the paper, we are currently in the process of anonymizing and preparing the video data for release. This involves removing any personally identifiable information and ensuring that all privacy concerns are addressed. We will make the video data available as soon as it has been processed and organized, so please stay tuned for updates.

Citation and License

Please cite this dataset as appropriate for your use. If you redistribute or use the dataset in downstream work, make sure your usage is compatible with the licenses of the underlying audio and spatial data sources.

@misc{zhu2026spatialomnispatialaudiounderstanding,
      title={Spatial-Omni: Spatial Audio Understanding Integration in Multimodal LLMs via FOA Encoding}, 
      author={Zhiyuan Zhu and Yixuan Chen and Yiwen Shao and Wenxiang Guo and Changhao Pan and Yu Zhang and Yuxiang Wang and Wei Liu and Houhua Zhang and Chengkuan Zeng and Wenbo Cheng and Yunxi Liu and Rui Yang and Steve Yves and Liefeng Bo and Zhou Zhao},
      year={2026},
      eprint={2606.10738},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2606.10738}, 
}