tnnanh1005/audioset-ver-1

AudioSet Emergency Sound Crops — v1 (preview)

1.5-second event crops mined from AudioSet for fine-tuning a 13-class emergency-sound model (EfficientSED fmn10+tf-256). Each clip is 16 kHz mono, 1.5 s, centred on the detected event.

v1 is a small preview/test subset (~20 clips per class), produced to validate the pipeline end-to-end. A full version will follow.

How it was built

1. Filter AudioSet by a 13-class emergency label map (+ hard-negative confusers).
2. Download the source 10 s clips and resample to 16 kHz mono.
3. Localise the event with the pretrained EfficientSED frame-wise model and cut a 1.5 s crop at the peak-activation frame (hard negatives cut at the energy peak).
4. Split by YouTube ID into train / validation / test.

Contents

• data/<split>/<role>_<ytid>.wav — the audio crops.
• metadata.csv — file_name, label, role, ytid, split.
  • role ∈ {positive, hard_neg}; for positives label is the final class.

Splits: train=353, validation=62, test=0 (total 415).

Positive classes

fire_smoke_alarm, emergency_siren, vehicle_horn, reversing_beeps, glass_break, gunshot, explosion, distress_scream, baby_cry, doorbell, door_knock, dog_bark, car_alarm.

Load

from datasets import load_dataset
ds = load_dataset("audiofolder", data_files="metadata.csv")  # or load_dataset(REPO_ID)

License & attribution

Derived from AudioSet (Google, CC-BY-4.0) via the agkphysics/AudioSet mirror and YouTube. Released under CC-BY-4.0. Localisation uses EfficientSED.