Titung/vehicle-audio-dataset

NINA: Naturalistic IN-vehicle Audio Dataset

Audio clips of sounds occurring inside (and, for emergency sirens, outside) a car cabin, rebuilt from axa-rev-research/NINA-Dataset following the original authors' recipe. Recordings come from dashcam/smartphone-mic YouTube compilations, so vehicle speed, device model, and microphone details are unknown (matches the source repo's own disclaimer).

⚠️ Provenance and usage notice

This dataset is derived from third-party YouTube videos (dashcam and news compilations), not originally-licensed recordings. The upstream repo ships a download + label recipe, not audio, likely for exactly this reason. By using this dataset you are responsible for complying with YouTube's Terms of Service and the copyright of the original video creators in your jurisdiction and use case. This repo is intended for non-commercial research (e.g. crash/anomaly sound detection). If you are a rights holder and want a clip removed, please open a discussion on this repo.

Classes

label	clips (this rebuild)
crash	751
driving	295
talking	265
horn	261
music	198
tire (tire skidding)	186
ambulance (siren)	160
scream (screaming)	157
pothole	144
meteo (strong rain/hail)	95
firetruck (siren)	76
police (siren)	39
door	34
acceleration (harsh acceleration)	22

Note: some videos referenced in the original youtube_IDs.csv may be unavailable (removed/private/ geo-blocked) at rebuild time, so exact counts can differ slightly run to run — see the notebook output above for this run's actual failed_downloads list.

Dataset structure

Standard Hugging Face audiofolder layout:

label_name/
  videoID_01-label_name.wav
  videoID_02-label_name.wav
  ...
metadata.csv   # file_name,label

Load with:

from datasets import load_dataset
ds = load_dataset("Titung/vehicle-audio-dataset")

How it was built

Rebuilt with yt-dlp (audio extraction) + ffmpeg (clipping) from the label files and video ID list in the upstream repo. See the companion notebook for the exact, reproducible pipeline.

Citation

If you use this dataset, please cite the original NINA dataset repo:

@misc{nina-dataset,
  title = {NINA: Naturalistic IN-vehicle Audio Dataset},
  author = {AXA REV Research},
  howpublished = {\url{https://github.com/axa-rev-research/NINA-Dataset}}
}

License

The upstream script is GPL-3.0. The audio itself is sourced from third-party YouTube content and is not covered by that license — see the provenance notice above.