humyn-labs/APAC-Egocentric-Residential-Voiceover

APAC Egocentric Residential (with Voiceover)

Ten narrated first-person recordings of household chores, each shipping the original capture with spoken voiceover, a burned-in caption render, WebVTT captions, and an ASS annotation track.

This is the only release in the HumynLabs egocentric collection that carries audio narration — the wearer describes each action as they perform it, and the captions align that speech to the video.

Preview: 45 s from the cooking sample, captioned render with audio. Turn sound on to hear the narration.

---

At a glance

Samples	10
Total duration	38 min 39 s
Audio	AAC stereo, 48 kHz — spoken narration on every sample
Caption cues	430 total
— narration	308
— sound events	122 (e.g. [egg cracking])
Resolution	mixed — 1920×1080, 2560×1440, 3840×2160, 3840×2880
Containers	7× MP4, 3× MOV
Files per sample	raw video · captioned render · .vtt · .ass

Samples

#	Sample	Category	Duration	Resolution	Cues
1	Cooking	Kitchen & Food Prep	6:35	1920×1080	71
2	Washing Dishes	Cleaning & Dishwashing	4:22	1920×1080	41
3	Folding Clothes	Laundry	4:12	1920×1080	62
4	Dough Kneading	Kitchen & Food Prep	4:36	1920×1080	55
5	Gardening	Outdoor	3:06	1920×1080	35
6	Cooking	Kitchen & Food Prep	3:26	3840×2160	25
7	Washing Clothes	Laundry	2:52	3840×2160	50
8	Shoe Cleaning	Household	3:25	3840×2160	34
9	Shopping	Retail	3:01	3840×2880	42
10	Peeling & Cutting	Kitchen & Food Prep	3:04	2560×1440	15

---

Repository layout

data/train-*.parquet   # 720p previews (with audio) + metadata + caption cues
videos/                # original captures — 7× .mp4, 3× .mov
captioned/             # renders with captions burned in
captions/*.vtt         # WebVTT, time-aligned
annotations/*.ass      # Advanced SubStation Alpha annotation track
preview/               # 720p proxies; card_sample_*.mp4 is the short clip above
metadata.csv           # flat table

The parquet embeds 720p previews with audio preserved, not the 7.85 GB of originals — the viewer stays fast and the narration is still audible. Originals are in videos/.

Columns

Column	Description
video	720p preview of the raw capture, audio intact — plays in the viewer
captioned_preview_video	720p preview of the burned-in caption render
sample_id	e.g. 01_cooking
sample, category	Chore and its category
duration_seconds, width, height, fps, codec	Probed from the media
has_audio, raw_container	True for all 10; mp4 or mov
raw_video_path, captioned_video_path	Full-resolution files in this repo
captions_vtt_path, annotations_ass_path	Caption and annotation sidecars
num_cues, num_narration, num_sound_events	Cue counts by kind
transcript	Narration cues concatenated — excludes bracketed sound events
cues	Every cue inline: start, end, text, kind
raw_sha256, raw_bytes	Integrity and size of the original

Caption format

WebVTT, with two kinds of cue interleaved:

00:00:02.860 --> 00:00:04.760 line:84%
Taking an egg and a fork.

00:00:07.797 --> 00:00:10.547 line:84%
<i>[egg cracking]</i>

00:00:08.420 --> 00:00:09.460 line:90%
Cracking the eggs open.

Spoken narration is plain text; sound events are bracketed and italicised. The kind field on each cue in cues separates them (narration vs sound_event), and transcript contains narration only — so you can train on speech without the sound tags leaking in.

Cues can overlap in time (the [egg cracking] cue above spans the narration cue that follows it). Treat the track as a set of intervals, not a partition.

The .ass files carry the same content with positioning and styling for rendering, which is what produced the captioned/ videos.

---

Usage

from datasets import load_dataset

ds = load_dataset("humyn-labs/APAC-Egocentric-Residential-Voiceover", split="train")
r = ds[0]
print(r["sample"], r["category"], r["num_cues"])
print(r["transcript"][:200])

Narration only, with timings:

narration = [c for c in ds[0]["cues"] if c["kind"] == "narration"]
for c in narration[:5]:
    print(c["start"], "→", c["end"], c["text"])

Originals, captions and annotation tracks:

from huggingface_hub import snapshot_download
snapshot_download("humyn-labs/APAC-Egocentric-Residential-Voiceover", repo_type="dataset",
                  allow_patterns=["videos/*", "captions/*", "annotations/*"])

Intended uses

Narrated activity understanding · egocentric video–language alignment · audio-visual ASR in noisy home environments · sound-event detection · procedural step segmentation · instruction-following for household robotics · caption generation and evaluation.

Limitations

• Small. 10 clips, 38 minutes — an evaluation and prototyping pack, not pretraining scale.
• Mixed resolution and aspect ratio. Samples span 1920×1080 to 3840×2880; sample 9 is 4:3, the rest are 16:9. Any fixed-size batching needs a resize and probably a letterbox step.
• Mixed containers — 3 samples are .mov, which some loaders handle differently from .mp4.
• Narration is scripted-in-the-moment, not conversational speech, and describes the action being performed — it is not a natural-dialogue ASR benchmark.
• No speaker labels, no word-level timings, and cue boundaries are phrase-level.
• Category skew: 4 of 10 samples are Kitchen & Food Prep; 2 samples are both "Cooking".
• No IMU, depth, hand pose, or action-segment labels — see the labeled monocular and stereo releases for those.

Provenance

Curated from the HumynLabs egocentric sample collection. All technical fields were probed from the media rather than copied from the source sheet — the sheet's DURATION column is formatted HH:MM:SS but the values are actually MM:SS. In the source sheet the link cells contain button labels ("Download · MP4", "Watch", "VTT", "ASS") rather than filenames; files were renamed deterministically as {index}_{sample}_{kind}.{ext}. All 40 files resolved and downloaded; no samples were dropped.

License

CC BY 4.0. Recorded with participant consent for research use. These recordings contain the wearer's voice as well as home interiors and incidental bystanders — please handle accordingly and do not attempt to identify individuals from voice or surroundings.

Citation

@misc{humynlabs2026apacvoiceover,
  title  = {APAC Egocentric Residential (with Voiceover)},
  author = {HumynLabs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/humyn-labs/APAC-Egocentric-Residential-Voiceover}
}