amberhandal24/acoustic-alert-5class

On-device Acoustic Alert 5-class Dataset

A curated mix of public acoustic-event datasets, mapped to 5 classes for real-time on-device alerting (Raspberry Pi-class hardware) without cloud calls or speech recognition.

id	label	alert?	description
0	background	no	Ordinary non-alert ambience: speech-like background, machines, footsteps, music, rain, etc.
1	dog_bark	yes	Dog barking and closely related dog vocalisations.
2	alarm_siren	yes	Sirens, alarms, fire alarms, vehicle sirens, alarm-like beeps.
3	door_event	yes	Door knocks and doorbells.
4	scream_shout	yes	Screams, shouting, yelling. Speech content is not transcribed.

Sources

source	role	license	upstream
ESC-50	small curated environmental sounds (dog, sirens, knocks, background)	CC BY-NC 4.0	https://github.com/karolpiczak/ESC-50
UrbanSound8K	additional dog barks, sirens, urban background	CC BY-NC 4.0	https://urbansounddataset.weebly.com/urbansound8k.html
FSD50K	larger Freesound-derived expansion (barks, alarms, knocks, screams, background)	CC0 / CC BY (per-clip)	https://zenodo.org/records/4060432
DogSpeak (3 000-clip subset)	extra dog barks (seeded dog_id-split subset)	per upstream HF repo	https://huggingface.co/datasets/ArlingtonCL2/DogSpeak_Dataset
Freesound (curated)	extra scream_shout and door_event clips, filtered to permissive licenses (publicdomain, zero, by/, sampling+)	mixed CC	https://freesound.org/

⚠️ AudioSet clips are intentionally not included in this dataset. AudioSet itself only distributes YouTube IDs (not audio) because the source material is copyrighted. The training mix used by the released pretrained model also pulled ~226 AudioSet clips locally; to reproduce that exact mix, fetch them yourself with the pick lists and scripts/fetch_audioset.sh in the project's GitHub repo (see "Reproducibility" below).

Splits and per-class counts (public mix, AudioSet excluded)

These reflect the exact rows in data/processed_5class_dogspeak/metadata.csv:

split	background	dog_bark	alarm_siren	door_event	scream_shout	total
train	18 409	~3 535	~1 861	~447	~585	~24 837
val	2 178	~514	~198	~78	~107	~3 075
test	3 831	~514	~601	~146	~325	~5 417

(DogSpeak is split by dog_id so the same dog never leaks across splits; ESC-50, UrbanSound8K, FSD50K use their official folds; Freesound is split via a deterministic MD5 hash of the clip id with 10 % val / 15 % test.)

What's in the repo

data/
├── raw/                               # source audio
│   ├── ESC-50-master/audio/*.wav
│   ├── UrbanSound8K/audio/fold{1..10}/*.wav
│   ├── FSD50K.dev_audio/*.wav, FSD50K.eval_audio/*.wav
│   ├── FSD50K.ground_truth/{dev,eval,vocabulary}.csv
│   ├── FSD50K.metadata/...
│   ├── DogSpeak/dogspeak_released/dog_*/...wav    (3 000-clip seeded subset)
│   └── freesound/{scream_shout,door_event}/*.wav
└── processed_5class_dogspeak/
    ├── metadata.csv          # per-clip label + train/val/test split
    ├── label_map.json        # class -> id mapping
    ├── dataset_summary.json
    ├── normalization.json    # feature params + train-split mean/std
    ├── feature_index.csv     # per-clip pointer to .npy feature
    └── features/*.npy        # Log-Mel features, shape (64, 64)

metadata.csv columns: filepath, label, label_id, split, duration, sample_rate, original_label, original_filepath, dataset, source_id.

Feature schema

Log-Mel spectrograms suitable for the released CNN baseline:

sample rate	16 000 Hz, mono
clip length	2 s (centre-cropped / zero-padded)
mel bins	64
FFT size	1024
hop length	512
range	dB-scaled, max-normalised, clipped at −80 dB
post-process	z-scored using train-split mean/std (see normalization.json)

To reproduce features from raw audio with the project's exact code, see scripts/extract_features.py in the GitHub repo.

Pretrained model

A lightweight CNN baseline (~28 k params) trained on this exact dataset:

• Test accuracy 0.74, macro-F1 0.55 (5 classes, held-out test split).
• Shipped with the GitHub repo under models/baseline_cnn_5class_dogspeak_v2/best_model.keras and runnable on a Raspberry Pi 4/5 with rpi/listen.py.

Reproducibility

Full pipeline (download → prepare → extract → train → eval → run on Pi) is in the GitHub project, branch amber-changes:

https://github.com/Zhu-Chenyu/On-device-Acoustic-Alert-System/tree/amber-changes

To rebuild this exact public mix from scratch:

git clone -b amber-changes https://github.com/Zhu-Chenyu/On-device-Acoustic-Alert-System.git
cd On-device-Acoustic-Alert-System
bash scripts/download_datasets.sh            # ESC-50 + UrbanSound8K + FSD50K + DogSpeak subset

# (optional) include Freesound clips: needs a Freesound API key
FREESOUND_API_KEY=xxx python scripts/fetch_freesound.py

python scripts/prepare_full_dataset.py \
  --fsd50k-root data/raw \
  --include-fsd-background --include-dogspeak --dogspeak-max 0 \
  --include-freesound \
  --exclude-labels baby_cry glass_break \
  --output-dir data/processed_5class_dogspeak
python scripts/extract_features.py \
  --metadata data/processed_5class_dogspeak/metadata.csv \
  --output-dir data/processed_5class_dogspeak

Privacy / intended use

This dataset is intended for acoustic event detection research and on-device assistive monitoring. The released model classifies event categories only; it does not transcribe speech and the inference pipeline saves no audio.

Citation

Please cite the underlying sources when using this dataset:

• ESC-50: Piczak, K. J. (2015). ESC: Dataset for Environmental Sound Classification. ACM MM.
• UrbanSound8K: Salamon, J., Jacoby, C., Bello, J. P. (2014). A Dataset and Taxonomy for Urban Sound Research. ACM MM.
• FSD50K: Fonseca, E., Favory, X., Pons, J., Font, F., Serra, X. (2022). FSD50K: An Open Dataset of Human-Labeled Sound Events. IEEE/ACM TASLP.
• DogSpeak: ArlingtonCL2 (2024). DogSpeak Dataset. Hugging Face.
• Freesound: https://freesound.org/ (please respect each clip's individual licence).

License

Combined: CC BY-NC 4.0 (driven by the most restrictive source licences, ESC-50 and UrbanSound8K). Per-source licences are listed above — respect each when using individual clips. Non-commercial use is required for the dataset as a whole.