Marine Acoustic Event Triage

Ship Transits (vessel sailing past the hydrophone)

Every closest-point-of-approach window is flagged as Ship Noise. Approach and recede tails produce Ambient — the physically correct answer because the ship is below the noise floor at those ranges.

Noise Resilience

The 50-step energy accumulation window averages short random transients. Sustained broadband contamination (storms, surf) above ~20 % degrades accuracy.

Sea-State Compensation

Whale Under Ship Noise

Fin-whale pulses interleaved with a simultaneous ship passage (2000-step mixed recording).

In the synthetic mixed scene, fin-whale pulses remain separately detectable during overlapping ship-noise windows. Real-ocean masking behaviour is pending MARS validation.

Remaining Limitations

1. Wider Gaussian Tuning — with_bandwidth(0.20)

Each Neuron::bandwidth is writable through the public API. v0.2 widens every channel past the library's auto-clamped 0.10 so the between-channel gaps no longer fall through to Ambient.

Bandwidth vs. CSV accuracy — sweet spot at 0.20

2. Multi-Label Output — step_multi()

Thresholds every channel independently and returns a MultiLabelDecision listing all active sources — so a fin whale calling through a simultaneous ship passage produces the label [FinWhale, ShipNoise].

Regression check: zero false-positive species on 2,000 steps of pure ambient ocean at bandwidth 0.20 — the multi-label path is strictly additive.

Usage

// v0.2: wider Gaussian tuning + multi-label output
let mut det = MarineDetector::with_bandwidth(0.20);

while let Some(d) = det.step_multi(freq) {
    // d.events may be [FinWhale, ShipNoise] simultaneously
    // d.energies == [FIN, BLUE, SHIP, HUMP] snapshot
}

Reproduce the full comparison locally: cargo run --release --example marine_v02.

Remaining Gap — 50 Hz

The 50 Hz window (between Fin=20 and Blue=80) is still reported as Ambient even at bandwidth 0.30 — the relative gap (80–20)/20 = 300% is too wide for any reasonable Gaussian to bridge. Species in this range (sei whale downsweeps 40–60 Hz) would need their own dedicated channel in a 5- or 6-channel bank.

RAM, ROM & Latency

CricketBrain is ~3× smaller than TFLite Micro's smallest audio example and ~5× smaller than a typical Edge Impulse audio model.

Average Power at 1 Hz Decision Rate

Edge Impulse doesn't publish mW numbers directly; estimates use the documented latency × ST datasheet power curves. At sub-1 mW budgets a solar-charged smart buoy can afford CricketBrain but not Edge Impulse.

Training Data & Accuracy Trade-off

On complex spectrograms (humpback song) CNNs beat CricketBrain on accuracy. On frequency-stable tonal signals (fin whale 20-Hz pulse, ship cavitation) the gap closes — and CricketBrain pulls ahead on every non-accuracy axis. Schall et al. (2024) note in their baleen-whale benchmark that deep-learning detectors "tend to be too computationally expensive to run on existing wildlife monitoring systems".

When to Pick Which

Full sourced breakdown (vendor docs, peer-reviewed papers): docs/competitive_analysis.md.