audio-alert-detector

Tiny CNN that classifies a 10-second audio clip as alert (siren, smoke/fire alarm, car alarm, house/burglar alarm) vs not-alert. Designed to run on a Raspberry Pi Zero 2 in the ONNX runtime with no accelerator.

Model

• Depthwise-separable CNN (~35k params, ~17M MACs)
• Input: raw mono 16 kHz PCM, 10 s = 160k samples
• Embedded log-mel frontend (no host-side preprocessing needed beyond resampling to 16kHz)
• Two heads: binary (alert/not) + auxiliary subclass (siren/alarm)
• Δ + ΔΔ time-derivative input channels for onset/sweep dynamics

Training data

• ~11k positives + ~60k negatives, 10 s each, 16 kHz mono
• AudioSet via confit/audioset-full HF mirror
• Targeted hard-negative mid lists (bells, whistles, woodwinds, mechanical, instruments, animals, music) for known FP categories
• Curated supplemental positives: ~199 country-specific EAS alarms + ~107 nuclear/civil-defense sirens

Training

• 40 epochs, AdamW (lr 3e-4, wd 1e-4), cosine LR schedule
• Batch 64, fixed 40% positive fraction per batch (uniform within each pool)
• Loss: binary BCE + 0.3 × subclass CE (masked to positives)

Augmentation (mel-space)

• Random time-stretch (0.9–1.1×)
• Random gain (–45 to +15 dB)
• Frequency shift (±4 mel bins ≈ ±2 semitones)
• Companding (γ 0.75–1.25, p=0.3)
• SpecAugment time + freq masks
• Curated ambience overlay (rain / cafe / road traffic / mic noise floor / etc.) at 25% RMS, applied to both classes

Deployment

Single-file fp32 ONNX (348 KB). Input: float32[batch, 160000] raw 16 kHz mono PCM. Outputs: binary_logit + subclass_logits[2]. Apply sigmoid for alert probability.

Recommended: 10s ring buffer, infer every 2 s, threshold 0.5, require 2 consecutive over-threshold windows to fire (eliminates almost all single-window FPs).

Performance

• Test set (225 curated clips): 88% acc / 89% prec / 83% rec / F1 0.86
• Pi Zero 2 (Cortex-A53, fp32, single thread): ~150 ms / 10 s window