Real-time passive acoustic monitoring

Listening to the ocean,
around the clock

A distributed network of underwater microphones, monitored continuously by machine learning. Detecting whale vocalizations as they happen, across two oceans.

Listen to a featured detection Loading...

Launch Dashboard

Streams Online

Total Detections

Public Detection Targets

Surfaced Detectors

About the project

Why passive acoustic monitoring matters

The ocean is vast, opaque, and noisy. Marine mammals communicate over distances of tens to hundreds of kilometers using sound — but human-generated noise, climate shifts, and habitat disruption are changing the acoustic landscape they depend on. Understanding who is vocalizing, where, and when is fundamental to conservation.

The challenge

Hydrophones generate continuous audio streams — thousands of hours per day across a distributed network. Manual review is impossible at scale. Vocalizations are transient, often buried in ambient noise, and span frequency ranges from 10 Hz infrasound to 100 kHz echolocation.

Our approach

Ocean Song ingests audio from 13 configured hydrophone sources across the Pacific and Atlantic. Two detector paths currently surface humpback and blue-whale candidates; other models remain in research or shadow evaluation. Available recordings include audio and spectrograms for listening and review. Source availability varies, and automated candidates are not individual scientific confirmations.

The network

13 hydrophones, 2 oceans

From the deep Monterey Canyon to the glacial waters of Disko Bay, we monitor available feeds around the clock. Stations can go offline; current connectivity appears in the dashboard.

MBARI MARS

Monterey Bay, CA — 900m depth
Detected: humpback, blue whale

Shoutcast

Disko Live

Disko Bay, Greenland
Detected: humpback

Shoutcast

Orcasound — Port Townsend

Admiralty Inlet, WA
Detected: humpback

HLS

Orcasound — Sunset Bay

Sunset Bay, WA
Detected: humpback

HLS

SIMRES

Saturna Island, BC
Detected: humpback

Icecast

Orcasound — North SJC

North San Juan Channel, WA
Detected: humpback

HLS

Orcasound — MaST Center

Des Moines, WA
Detected: humpback

HLS

Orcasound — Andrews Bay

Andrews Bay, WA
Detected: humpback

HLS

Orcasound — Point Robinson

Vashon Island, South Puget Sound, WA
Listening — no detections yet

HLS

Orcasound Lab — Haro Strait

West San Juan Island, Haro Strait, WA
Listening — no detections yet

HLS

Bush Point

Whidbey Island, Admiralty Inlet, WA
Listening — no detections yet

HLS

Lime Kiln

Lime Kiln Point, San Juan Island, WA
Listening — no detections yet

Shoutcast

ONC Barkley Canyon (NEPTUNE)

Off Vancouver Island, BC — ~640m slope
Listening — no detections yet

ONC

How it works

From hydrophone to detection

🌊

Hydrophones

Underwater microphones streaming 24/7 from coastal and deep-ocean stations

🔄

Ingestion

Real-time audio capture at 16-48 kHz via HLS, Icecast, Shoutcast, and ONC protocols

🧠

Detection

2 surfaced detection models analyzing every 10 seconds: TF Hub, PyTorch, and DSP classifiers

🎵

Recording

Automatic clip extraction, spectrogram generation, SNR measurement, denoising

Species detection

What we listen for

Humpback Whale

Megaptera novaeangliae

Model: Google TF Hub humpback_whale
Range: 80 – 4,000 Hz
Method: Mel spectrogram CNN, 3.9s windows
Threshold: 90%, 2 consecutive

Orca — not currently surfaced

Orcinus orca

Model: OrcaHello (orcasound), running in shadow
Range: 500 – 10,000 Hz
Status: detections withheld since 2026-07-28. A 60-clip audit measured 10.9% precision (95% CI 4.7–23%); the clips were mostly vessel and impulsive noise, and precision did not improve at higher confidence. Scoring continues for recalibration.

Blue Whale

Balaenoptera musculus

Model: MBARI ResNet50 A-call
Range: 10 – 100 Hz
Method: Spectrogram classifier
Status: Input preprocessing under review; field accuracy not established.
Threshold: 70%

Not surfaced

shadow & disabled detectors

Sperm whale, fin whale, and dolphin detectors are not surfaced publicly. Their current models produce too many false positives — e.g. boat-knocking mimics sperm-whale clicks — so sperm-whale runs in shadow (scoring only) and fin/dolphin are disabled until validated detectors replace them.

Field notes

Lifting humpback song from the noise floor

We trained a humpback-specific denoiser that lifts song out of vessel noise, flow, and deep-sea hum. See and hear before/after recordings from hydrophones around the world — the first in a series on the acoustics and machine learning behind the project.

Read the field note

Explore the live dashboard

Browse real-time hydrophone streams, listen to recent detections, and denoise whale vocalizations.

Enter Dashboard

Ocean Song / bioacoustic monitoring

0 sources

0 monitored

0 detections

System: offline

Analysis Pipeline

Google Humpback Whale TF Hub · tfhub.dev/google/humpback_whale/1 · Threshold: 90% · 2 consecutive windows

Humpback whale detection · All 13 streams · 10kHz input · 3.9s windows

OrcaHello orcasound/aifororcas · ResNet50 (SRKW) · SHADOW — withheld, 10.9% measured precision

Orca detection · 9 Salish-Sea streams (Orcasound + Bush Point + Lime Kiln) + MBARI · three-condition gate: OrcaHello-high, humpback-low, and signal-present (suppresses song FPs + near-silent peak-norm hallucinations)

MBARI Blue Whale A-call MBARI · ResNet50 classifier · Threshold: 70%

Blue whale A-call detection · MBARI MARS stream · 2kHz input · Preprocessing under review; field accuracy not established

Sperm Whale Click Train Shadow — scores logged, not surfaced (boat-knocking false positives)

Fin Whale · Dolphin Disabled — produced false positives, pending better models

13 streams monitored · 16kHz ingestion · 5-min ring buffers · 3 active models · Inference every 10s

Reference Archive

Curated, openly-licensed humpback recordings, scored with the same metrics as our live detections. Pick one on the left to hear it; open it in Listen to denoise & analyze. All credit their source.

Loading archive…

Review & Label

Help assess our research recordings: listen to each clip, inspect its spectrogram, and record what you can identify. Model metrics and hypotheses are prompts for review, not confirmed labels. Clips may contain noise, silence, or uncertain species. Your notes support evaluation; a trusted-code flag does not establish expert or independent validation.

Loading review queue…

Research Dataset

The dataset our labeling loop builds — three tracks below (clean song, noise library, species finds). Pick a clip on the left to hear it. Grows as the loop runs.

Loading dataset…

Tuning in…

Selected Event

Select an event from the sidebar

Denoise

Selected Event

Select an event from the sidebar

No event selected

Source Separation

—

Agent & API Access

Base URL: https://api.oceansong.live

Swagger Docs agent.json Claude Code Skill OpenAPI JSON

Quick start:
curl https://api.oceansong.live/summary
curl https://api.oceansong.live/highlights?hours=24

No authentication required · WebSocket at wss://api.oceansong.live/ws

Questions, feedback, collabs? Telegram @roman_still →

Listening to the ocean,around the clock

Why passive acoustic monitoring matters

The challenge

Our approach

13 hydrophones, 2 oceans

From hydrophone to detection

Hydrophones

Ingestion

Detection

Recording

What we listen for

Lifting humpback song from the noise floor

Explore the live dashboard

Listening to the ocean,
around the clock