Acoustic monitoring for bird research

Hear every bird
in the chorus.

We build field-ready microphone arrays and software that pull individual birds out of the dawn chorus, work out where each one is singing, and identify it right on the device.

See how it works Follow our progress

The problem

One microphone hears a crowd.

At dawn in the tropics, dozens of species sing at the same time. Most recorders listen through a single microphone, so overlapping songs blur together and AI classifiers miss the quieter voices.

Microphone arrays that can separate voices already exist, but they're tuned for people in meeting rooms. Songbirds sing much higher than we speak, often above the range those arrays were built for.

We're building the array, the separation model, and the software for birds.

How it works

From the canopy to your screen, in four steps.

  1. 01

    Listen

    Eight microphones record the forest together. A song reaches each one a fraction of a millisecond apart, and that tiny difference gives away its direction.

  2. 02

    Separate

    Our model finds each singer's bearing, then pulls its voice out of the mix. Each bird gets its own clean recording, tagged with the direction it came from.

  3. 03

    Identify

    Google's Perch 2.0 model names the species on the device, narrowed to what's likely for your location and season. A local model learns from the calls you confirm.

  4. 04

    Explore

    A live dashboard shows who's singing, when, and from where. It runs on the device in the field and in the cloud when you're back home.

The Oscine Acoustics prototype: a weatherproof enclosure topped by a round sensor board with eight microphones in a ring
First field prototype

01 · The array

A microphone array built for birdsong.

Eight MEMS microphones sit in a small ring, spaced so the array can tell apart birds singing from different directions, even at songbird pitch.

  • Tuned for birds. Optimised for 500 Hz to 12 kHz, where most passerines sing.
  • Low noise, high sensitivity. Quiet, distant calls still come through clearly.
  • Built to stay outside. A resin-filled housing keeps out humidity, heat and tropical rain.
  • Edge-ready. Plugs into a Raspberry Pi 5 today, with an NPU-powered board on the way.

02 · Source separation

Direction first, then detail.

Many separation models are asked to untangle a recording on their own. In real forests they tend to give up and simply turn the volume down. We take a different route.

  1. 1
    Find the singers. Classic acoustic maths scans all 360° and picks out the directions where birds are calling.
  2. 2
    Clean up each voice. A compact neural network refines one direction at a time, keeping the bird and leaving the rest behind.
  3. 3
    Hand off clean audio. Up to four birds come out as separate recordings, each with its bearing, ready for identification.

Because the network starts from the classical result, it can only improve on it. It is trained on real recordings from our field site and small enough to run on the device.

Mixture
Bird 1 · 51° NE
Bird 2 · 247° WSW
Bird 3 · 175° S

03 · Edge & cloud

A field station that runs itself.

Each station records, separates and identifies birds on its own, with no internet required. When a connection is available, it syncs to the cloud so you can follow every site from anywhere.

Station 01 · Sierra Nevada de Santa Marta Live

Recent detections ▶ tap to listen

    Where they're singing

    Activity today ☀ sunrise 05:52

    Real detections from our Santa Marta station, with the direction each bird sang from. Click one to listen.

    Works offline

    Recording, separation and identification all happen on the station. Weeks without signal are fine.

    Learns your local birds

    Confirm or correct a detection and a local model learns from it, picking up regional dialects and ignoring familiar noise.

    Location-aware

    Suggestions are filtered by where the station is and the week of the year, so unlikely species don't slip in.

    Review faster

    Unknown calls are grouped by similarity, so you can label a whole cluster at once instead of clip by clip.

    Patterns at a glance

    Hourly heatmaps with sunrise, sunset and weather, plus compass charts of where each species sings.

    One dashboard, everywhere

    The same interface runs on the station and in the cloud, with secure live listening and multi-site accounts.

    In the field

    Tested where the chorus is loudest.

    Our prototype lives in a remote corner of the Sierra Nevada de Santa Marta in northern Colombia, one of the most bird-rich mountain ranges on Earth. It faces heavy rain, high humidity and strong sun every day.

    Dawn there brings dozens of species singing at once. That makes it a tough place to work and the right place to train and test our separation model on real recordings.

    A rainbow over green Andean foothills, seen from the field station

    Where we are

    1. Prototype array deployed

      Recording daily in the Santa Marta mountains.

    2. Edge & cloud platform

      Live detections, review tools and cloud sync running on Raspberry Pi 5.

    3. Separation model

      Trained on real field recordings and being tuned to run in real time.

    4. Next-gen hardware

      A board with a neural accelerator for full-quality separation on the device.

    5. Research availability

      Projected for late 2027.

    Stay in the loop

    Follow the project.

    We share field notes, results and hardware news a few times a year. Researchers and conservation teams interested in early deployments are welcome to leave a note.