ShantyCam/audiodet-synth-smoke

AudioDet — Synthetic Smoke-Alarm Clips

250 fully synthetic smoke-alarm audio clips generated procedurally with NumPy. No third-party audio was used — every sample is a parametric tone pattern written by gen_smoke_alarms.py. This makes the set cleanly licensable (CC-BY-4.0) and deterministically reproducible.

They were created as clean positive examples of the class smoke_alarm for the ShantyCam AudioDet sound-event detector, so a model can learn the alarm cadence without needing scarce, restrictively-licensed real recordings.

What's in it

• data/smoke_alarm_synth_000.wav … smoke_alarm_synth_249.wav
• 16 kHz mono, float32 WAV, 8 s each

Signal model

The T3 temporal-three pattern used by residential smoke alarms:

3 × (beep + gap) + long pause,  repeated for 8 s

Randomized per clip:

• Timbre: ~3.1 kHz sine (modern piezo) or ~520 Hz band-limited square (bedroom standard)
• Detuning: ±7 % around the base frequency
• Level: −30 … −8 dBFS
• Beep / gap: 0.45–0.60 s each; pause: 0.9–1.6 s
• AM tremor: optional 4–9 Hz, 8 % depth (piezo imperfection)

Clips are clean (no background). Mixing into an acoustic environment (reverb + background noise) is left to the training-time augmentation stage.

Reproduce

pip install numpy soundfile
python gen_smoke_alarms.py   # writes 250 clips (seed = 1234)

Intended use

Positive-class training data for smoke-alarm / sound-event detection. Because the clips are idealized synthetic patterns, pair them with real background/negative audio and augmentation for a deployable model.

License

CC-BY-4.0. Attribution: ShantyCam — AudioDet Synthetic Smoke-Alarm Clips.

Sources / references (NOT redistributed here)

This dataset contains only ShantyCam-generated synthetic audio. It does not include or redistribute any third-party corpus. The broader AudioDet model was trained against public datasets that remain under their own licenses and are referenced, not re-hosted: AudioSet, FSD50K, LibriSpeech, MUSAN, OpenSLR RIRS_NOISES, Clotho, Barkopedia (permissive/foreign) and ESC-50, UrbanSound8K, TUT Rare Sound Events 2017, CryCeleb (non-commercial / restricted — reference only). See the model repo's LICENSING.md for the full provenance audit.