Sphere360

Paired first-order ambisonic audio and 360° equirectangular video, sourced from YouTube. Each pair shares one ID, so the spatial sound field and the visual sphere are aligned by construction. 791 hours of paired material.

	Pairs	Audio	Video	Size
train	5,469	5,469	5,469	5.24 TB
test	162	162	162	0.17 TB
extras	—	415	110	0.25 TB
Total	5,631	6,046	5,741	5.66 TB

Audio — Opus, 48 kHz, 4-channel first-order ambisonics (ambisonic 1). Every audio file in the dataset is verified to carry exactly this layout. Video — VP9 equirectangular. It also carries a stereo track that is not the spatial audio; always take the ambisonics from audio/.

Layout

audio/{train,test}/{00..63}/<id>.webm     ambisonic audio
video/{train,test}/{00..63}/<id>.webm     360° video
extras/audio_only/{train,test}/<id>.webm  unpaired
extras/video_only/{train,test}/<id>.webm  unpaired
metadata/{train,test,extras}.jsonl        one row per ID
metadata/stats.json                       counts, and every file excluded and why

The 64 shards keep each directory to ~184 files and leave room to grow. Shard is derived from the ID, so no lookup is needed:

import hashlib
shard = lambda vid: f"{int(hashlib.sha1(vid.encode()).hexdigest()[:8], 16) % 64:02d}"
shard("-EOhDHns4xw")  # -> "00"

Shards are hashed rather than cut from the ID's first characters because YouTube IDs are case-sensitive and would collide on case-insensitive filesystems.

Use

Extensions vary, so read paths from the metadata rather than building them:

import json
from huggingface_hub import hf_hub_download

R = "OmniAV/Sphere360"
rows = [json.loads(l) for l in open(hf_hub_download(R, "metadata/train.jsonl", repo_type="dataset"))]

r = rows[0]
audio = hf_hub_download(R, r["audio"]["path"], repo_type="dataset")
video = hf_hub_download(R, r["video"]["path"], repo_type="dataset")

One row:

{"id": "-EOhDHns4xw", "split": "train", "shard": "00", "paired": true,
 "youtube_url": "https://www.youtube.com/watch?v=-EOhDHns4xw",
 "audio": {"path": "audio/train/00/-EOhDHns4xw.webm", "bytes": 11536470,
           "duration": 255.981, "codec": "opus", "channels": 4,
           "channel_layout": "ambisonic 1", "sample_rate": 48000},
 "video": {"path": "video/train/00/-EOhDHns4xw.webm", "bytes": 557769482,
           "duration": 256.008, "codec": "vp9", "width": 3840, "height": 2160}}

Pull one shard instead of 5.66 TB:

hf download OmniAV/Sphere360 --repo-type dataset --include "audio/train/00/*" "video/train/00/*"

Know before you train

• Resolution is not uniform — 56 distinct sizes. 3840×2160 (44%), 3840×1920 (28%) and 3840×2048 (9%) cover most of it, but the range runs from 1920×960 to 7680×4320. This reflects the sources, not a processing error. Filter on width/height if your pipeline needs one size.
• Duration is not uniform — 12 s to 11.4 h, median 277 s. Clip as needed.
• extras/ is unpaired and excluded from train/test. Single-modality use only.
• 30 files were dropped, each a low-bitrate video rendition the fetcher had written into the audio slot when the ambisonic track was unavailable. They contained no audio stream at all. Their IDs are listed in metadata/stats.json under rejected_files; the 27 that still had real video moved to extras/video_only/.
• Material was fetched twice in parallel. Where both runs returned an ID, the higher-resolution copy won, with file size breaking ties at equal resolution.
• Every row carries youtube_url, so any file can be traced to its source.

Availability

The dataset is uploaded in shards. metadata/*.jsonl lists only rows whose files are already present, and metadata/stats.json records the published shards under published_shards — so a run driven by the metadata never asks for a missing file. Counts in the table above are the full target; check published_rows for what is live right now.

Citation

Derived from the Sphere360 dataset introduced in OmniAudio.

@inproceedings{omniaudio2025,
  title     = {OmniAudio: Generating Spatial Audio from 360-Degree Video},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2025}
}