Skip to content

OCEAN SENTINEL / ACOUSTIC MODEL · V7.6

From sound.To detection.

A 2.4M-parameter convolutional network trained on 197,000 labelled audio windows across 29 hydrophones. Per-site calibration turns model confidence into an inspectable vessel detection.

Explore the architectureCLI & deployment
SIGNAL → FEATURES / CONCEPT

MODEL / V7.6

Evaluation & deployment profile
01Held-out test accuracy · v7.6 + per-site calibration
96.4%
02Out-of-distribution accuracy · 8 sites × fresh dates
96.0%
03End-to-end inference latency · M-series MPS
19 ms
04Hydrophone sites in the calibration set
29

SYSTEM ARCHITECTURE / 08 OPERATIONS

One recording.
An inspectable decision.

Eight operations. One traceable path from hydrophone to operator. Scroll through the pipeline and inspect how sound becomes a decision you can act on.

FOLLOW THE SIGNALSCROLL TO EXPLORE 01 / 08
ACQUISITION / 01
00 sAMPLITUDE / TIME60 s
Time domainIllustrative schematic · not live inference

OPERATION 01 / ACQUISITION

Ingest audio

Clean 60-second window

The CLI resamples each clip to 16 kHz mono, then pads or crops it into one precise minute so every deployment starts from the same signal.

INPUT
Hydrophone / WAV
OUTPUT
16 kHz · mono · 60 s

The signal changes form.
The evidence stays with it.

Run the pipeline
MODEL → DEPLOYMENT01 / INSTALL

01 / INSTALL

One package.
The complete stack.

The checkpoint, calibration tables, sample clips and CLI ship together. Install Ocean Sentinel, then verify the system before processing your first recording.

Install package
pip install ocean-sentinel
Verify installation
os doctor

Loads the v7.6 weights, runs a synthetic forward pass and checks each component. A non-zero exit code makes failures visible in CI.

  1. 01

    Model checkpoint

    CNN v7.6 weights

  2. 02

    Site calibration

    Local decision thresholds

  3. 03

    Sample audio

    Clips to test the pipeline

  4. 04

    Operator tools

    Every CLI subcommand

Inspect the CLI exampleCOMMAND + OUTPUT
Ocean Sentinel CLI
MacBook Pro ~ % pip install ocean-sentinel
 Collecting ocean-sentinel
 Installing collected packages: ocean-sentinel
 Successfully installed ocean-sentinel-7.6.0

MacBook Pro ~ % os doctor

 health check · v7.6 · mps (Apple Silicon)

 ✓ model cnn_v7_6.pt 9.1 MB · sha 4f2c…
 ✓ thresholds 10 sites per_site_thresholds_v7_6.json
 ✓ conformal α = 0.05 · threshold p ≥ 0.480
 ✓ pytorch 2.11.0 device mps
 ✓ librosa 0.11.0 forward pass OK
 ✓ e2e inference ship_prob=0.280 latency 19 ms
 ✓ eval results 2 / 2 files present

 All checks passed. System is healthy.

MacBook Pro ~ %

Example output from the CLI walkthrough. No command is executed by this page.

LOCAL SOUND → LOCAL THRESHOLD02 / CONFIGURE

02 / CONFIGURE

Every site sounds
different.

Gemma guides an eight-step onboarding protocol, from ambient recording to a saved site profile. Fourteen tools handle reference matching, adaptation, calibration and alert policy.

Explore onboarding
os onboard --demo

A guided workflow of about five minutes, with the hydrophone’s own false-alarm budget at its center.

HYDROPHONE / SITE PROFILE8 STEPS · 14 TOOLS
SITE ACOUSTICSPROFILE / 01
Acoustic profile concept · four phases of the eight-step protocol

Start with the site.

Record the local ambient sound and look up the AIS baseline. Establish the acoustic conditions before setting an alert boundary.

OUTPUTAmbient recording + AIS baseline
Inspect the CLI exampleCOMMAND + OUTPUT
Ocean Sentinel CLI
MacBook Pro ~ % os onboard --demo

●●●○○○○○ Step 3 of 8 Transfer learning · find nearest known sites
 ────────────────────────────────────────────────────────

—compare_to_known_sites
 signature = "<vec[64]>"
 nearest training sitesMBARI MARScosine0.84 ✓ recommended
monterey-bay-aqcosine0.71
port-townsendcosine0.43

●●●●○○○○ Step 4 of 8 Per-site adapter · fine-tune last 2 layers
 ────────────────────────────────────────────────────────

—finetune_adapter site=monterey-test · 10 epochs · lr=3e-4
 epoch 8/10 loss 0.700 val_acc 0.844
 epoch 9/10 loss 0.610 val_acc 0.872
 epoch 10/10 loss 0.520 val_acc 0.900

 ✓ adapter trained · val_acc 0.892 · saved adapter.pt

Example output from the CLI walkthrough. No command is executed by this page.

AUDIO → EXPLAINED EVENT03 / DETECT

03 / DETECT

One file.
Every reason.

Run the full stack on a recording. Audio becomes a 128-band spectrum, the CNN scores the signal, and the site profile sets the decision boundary.

Detect a recording
os detect data/clip.wav --site olympic-pilot --ais 0
--site
Apply the hydrophone’s calibrated threshold.
--ais N
Add the nearby vessel count to determine the alert tier.
--json
Return the same detection as structured output.
DET-57313ILLUSTRATIVE CLI RECORD
ACOUSTIC DETECTIONDark vessel.

olympic-pilot / clip.wav

CNN CONFIDENCE0.927
0Site boundary / 0.5201
Uncertainty
0.134 / abstain > 0.25
AIS context
0 vessels / 10 km radius
Alert tier
DARK_VESSEL / HIGH
GEMMA / EXAMPLE NOTE“Strong ship signature, zero AIS in radius. Conformal threshold passed comfortably.”
WHY THIS ALERT

Acoustic evidence crosses the local boundary. Maritime context determines the tier.

Inspect the CLI exampleCOMMAND + OUTPUT
Ocean Sentinel CLI
MacBook Pro ~ % os detect data/clip.wav --site olympic-pilot --ais 0

—run_pipeline
 clip = "data/clip.wav"
 site = "olympic-pilot"
 pipeline traceCNN confidence0.927(ship)
conformal passyes(p ≥ 0.520)
uncertainty0.134(abstain > 0.25)
AIS context0 vessels in 10 km
decision tierDARK_VESSEL
severityHIGH
latency597 msload + mel + CNN

—explain_decision id = "DET-57313"
 GemmaStrong ship signature, zero AIS in radius.
Conformal threshold passed comfortably.
If this were live, I'd page you immediately.

Example output from the CLI walkthrough. No command is executed by this page.

CONTINUOUS INPUT → OPERATOR VIEW04 / MONITOR

04 / MONITOR

Keep listening.
Keep the evidence.

Point the watcher at incoming WAV files. Each stable recording moves through the same detection pipeline, with its explanation and visual evidence attached.

Monitor incoming audio
os monitor olympic-pilot --watch ./incoming/
  1. 01 / WATCH

    Wait for a complete file.

    A three-second stability check protects against partial writes.

  2. 02 / RECORD

    Keep the local record.

    Append detections to events.jsonl and send them to the ingest gateway.

  3. 03 / DELIVER

    Bring the evidence.

    Spectrum, Grad-CAM and analyst note reach the dashboard with the event.

ILLUSTRATIVE SESSION / OLYMPIC-PILOT4 EVENTS / 3 MIN
DET-57313

Dark vessel

DARK_VESSEL / HIGH

Strong acoustic evidence with no nearby AIS. The example crosses the site threshold of 0.52.

Pushed to gateway · spectrum + saliency
SELECTED WINDOW / clip_1759.wavExample data · no live connection
Inspect the CLI exampleCOMMAND + OUTPUT
Ocean Sentinel CLI
MacBook Pro ~ % os monitor olympic-pilot --watch ./incoming/

—watch poll 2 s · stable 3 s · gateway live

14:02:08 clip_1759.wav ● DARK_VESSEL HIGH
 p = 0.93 > 0.52 id DET-57313
 ↳ pushed to gateway · spec + saliency uploaded

14:03:12 clip_1760.wav ○ AMBIENT
 p = 0.08 ≤ 0.52

14:04:18 clip_1761.wav ● ACOUSTIC_ONLY_LOW MEDIUM
 p = 0.71 flagged for review
 ↳ pushed to gateway

14:05:24 clip_1762.wav ● GONE_DARK_VESSEL HIGH
 p = 0.91 AIS broadcast just stopped
 ↳ pushed to gateway

monitor stopped — 4 events in 3 min

Example output from the CLI walkthrough. No command is executed by this page.

FROM MODEL TO FIELD

The next signal.
A clearer picture.

Follow a detection into the operator view, or trace the OC01 investigation and its subsequent audit.