Hear every reserve.

An open acoustic monitoring system. Solar stations listen to birds, rangers see everything on one map.

Join the pilot How it works

Reserves are vast. People in the field are few. Birds do not wait for the next survey.

At night

Many birds fly and call after dark, when nobody is there to hear.

In migration season

Waves of birds pass in days. Rare visits miss them.

Far from people

A huge territory and few people in the field.

A microphone listens around the clock. And it is cheap.

It listens.

It understands.

It tells you.

How it works

Only the event leaves the device.

A MEMS microphone on a solar-powered ESP32-S3 station records day and night.

A small model names the species on the device itself. The raw audio stays there.

LoRa carries the species, the time and the confidence. A tiny message, not a recording.

A gateway in the reserve office receives the events. No internet needed.

Rangers get an event feed, a map of stations and Telegram alerts.

Two parts. One system.

A station in the field and a platform in the office.

Prototype

Station

A low-cost, solar-powered listening device. It runs the species model itself and sends only events.

  • ESP32-S3 with an I2S MEMS microphone
  • LoRa radio for remote terrain
  • Solar panel and battery
  • On-device species model
In design

Platform

One screen for the whole reserve. Works offline with a local gateway.

Flamingo05:42high
Demoiselle crane06:10medium
Lapwing06:31high
TelegramPossible gunshotStation 3
TelegramEngine noiseStation 3
CSV
Darwin Core
GBIF

Event feed with an audio clip to verify

Planned

Open by design

Data and code will be published, so any reserve can repeat the setup.

See what rangers see.

A live preview with demo data. Real events will look like this.

Demo data

Event feed

    Stations

    Architecture

    Five parts, one database.

    Four sources write to the database. The dashboard reads only from it.

    • Bioacoustic AINow

      Sound from the stations. The core of the project.

    • Wetland Remote SensingIn parallel

      Water and vegetation from open satellite imagery.

    • AI Bird VisionSecond stage

      Photos and headcounts to calibrate the sound index.

    • Migration IntelligenceLater

      Partners' GPS tracks and observations.

    One database

    observations: species, time, place

    environment layers: water, vegetation

    National Biodiversity Dashboard

    • Map of Kazakhstan
    • Abundance dynamics
    • Migration corridors
    • Anomaly alerts

    Built for people who protect the land.

    Reserves

    Know what lives on your territory around the clock. Get alerts about rare species and threats. Keep your own data.

    Scientists

    Labeled recordings from the Kazakh steppe, BirdNET and Perch tested on local species, export in Darwin Core.

    Rangers

    Alerts in Telegram, a map of stations, nothing complicated to set up.

    First site: Korgalzhyn.

    A wetland of international importance in the Kazakh steppe.

    UNESCO

    Part of the World Heritage site Saryarka, Steppe and Lakes of Northern Kazakhstan (2008).

    Ramsar

    A wetland of international importance: the Tengiz and Korgalzhyn lake system.

    Flamingos

    Lake Tengiz is described as the northernmost flamingo nesting site in the world.

    Sources: UNESCO (whc.unesco.org), Ramsar Convention (ramsar.org)

    From one reserve to a whole country.

    Events from every reserve flow into one national dashboard in Astana. More reserves join the same system.

    • Pilot: Korgalzhyn
    • Next reserves
    • National dashboard
    Astana Korgalzhyn

    Validation

    How we test it.

    Baseline

    BirdNET and Perch on existing recordings give a first accuracy table by species.

    Own dataset

    Korgalzhyn recordings labeled with the reserve's ornithologist, so the accuracy is verified.

    Honest split

    We test on other sites and other dates, not on random clips of one recording.

    Field test

    A reference recorder runs alongside. Wind and water are tested separately. False alarms are counted per hour.

    Six steps. Each one shows a result.

    Contact

    The reserve agrees to the installation and names an advising ornithologist.

    Baseline

    First accuracy table for BirdNET and Perch.

    Dataset

    Collection and labeling, split by site and date.

    Model and MVP

    Fine-tuning, quantization, event feed and Telegram alerts on one station.

    Field season

    Power draw, LoRa range, comparison with a reference recorder.

    Analysis and defense

    Results for the RKNP and ISEF science competitions.

    In parallel: satellite analysis of water and vegetation.

    Next: a second reserve with different terrain, to test whether the model transfers.

    Be the first reserve to listen.

    We are looking for a pilot reserve, an ornithologist, and mentors in audio ML and remote sensing.

    Join the pilot