01 / RECORD
Record
Recorders, the partner's own or ours, are placed across the site. Each records a minute of audio every five minutes to a memory card. They only record; no analysis runs on the device.
+20%
BIRD DETECTION SCORE VS BIRDNET 3.0, WITH SITE TRAINING
11 vs 2
SPECIES IN 30 MINUTES VS AN OFF-THE-SHELF PIPELINE
13.5×
FEWER DETECTIONS: A LIKELY FAULTY RECORDER, SPOTTED FROM ITS DATA
6 h
OF AUDIO FOR A FIRST PROFILE OF A SITE
HELD-OUT TEST SET · RESTORATION PARTNER DEPLOYMENT, SOUTH INDIA · BERI FIELD TEST. DETAILS BELOW.
[ 01 · THE GAP ]
Trees planted is not the same as a forest recovering.
Restoration programmes, reserves and CSR projects are measured by saplings in the ground. Whether birds, frogs and insects come back, or whether someone is cutting at the edge at night, is hard to know without surveys that happen once a season.
Sound fills that gap. A recorder listens day and night, in places a person cannot stay. The hard part is the hours of audio that come back, far more than anyone can listen to. That is the part Canopy does.
[ 02 · HOW IT WORKS ]
Simple hardware in the field.
The heavy work happens in the cloud.
02 / COLLECT
Collect
After a deployment the memory cards are collected and the audio comes to us. A speech check flags minutes that may contain people talking: nobody listens to them, and they are never shared.
03 / ANALYSE
Analyse
Our bird model names the species, filtered to the ones that occur in the region. A second model, now in testing, is set to the site's own background and listens for chainsaws and gunshots, and anything else out of place goes to a list for a person to review.
04 / REPORT
Report
A plain report: which species were heard and how sure we are, day versus night, and every flagged event with the audio to check. Repeat deployments will show change over time.
[ 03 · THE MODEL ]
Tested against the standard models.
We benchmark Canopy's bird model against the models acoustic projects commonly use: BirdNET 3.0, BirdNET 2.4 and Google's Perch 2.0.
In the lab, on 21,781 clean public recordings and with no site training, it beats each of them. In real field audio, where birds are distant and overlapping, and trained on a site's checked recordings, it scores 20% higher than BirdNET 3.0 on recorders it never saw in training.
Threats go through a separate detector, now in testing, whose alert line is set from each site's own background and never drops below a floor. Replayed over our Beri recording, where nothing threatening happened but an engine ran throughout, it raised about 1.6 false alerts a day.
BENCHMARKS · MACRO AVERAGE PRECISION, 1.0 IS PERFECT
LAB · CLEAN CLIPS · NO SITE TRAININGVS BN 3.0
Canopy0.771+2%
BirdNET 3.00.757
BirdNET 2.40.731
Perch 2.00.604
REAL WORLD · FIELD AUDIO · SITE-TRAINEDVS BN 3.0
Canopy0.486+20%
BirdNET 3.00.405
Perch 2.00.379
BirdNET 2.40.314
SIMULATED · THREAT DETECTOR AT BERI
Chainsaws flagged at 100 m79%
False alerts per day at a noisy roadside site~1.6
LAB: BIRDCLEF 2024, 21,781 PUBLIC RECORDINGS, EACH FOCUSED ON ONE BIRD. REAL WORLD: 1,152 RAINFOREST CLIPS FROM 6 RECORDERS HELD OUT OF TRAINING, 68 SPECIES; MANY DISTANT, OVERLAPPING BIRDS, SO EVERY MODEL SCORES LOWER THAN IN THE LAB. COMPARE WITHIN A TEST, NOT ACROSS. SIMULATED: RECORDED CHAINSAWS MIXED INTO REAL BERI BACKGROUND, NOT A LIVE INCIDENT.
[ 04 · PARTNER DEPLOYMENT ]
A lake restoration site in South India.
Our restoration partner ran four recorders, within about 15 m of each other, at a lake for eleven days: about 214 hours of audio. They shared the recorder logs, their existing pipeline's detections and a sample of the audio, and asked us what it all showed.
In 30 minutes of audio from one recorder, we named 11 species where their pipeline, which keeps only detections scored 0.8 or higher, listed 2. We removed 26 of the 66 species it had reported across the site, because none has a recorded occurrence in India. We also spotted a recorder that was probably faulty from its data alone: it logged 13.5 times fewer detections than the other three, and the fewest on every one of the eleven days.
Minutes that might contain people talking were set aside without anyone listening to them. Our report went to the partner first, and they agreed to these results being shown without names.
DEPLOYMENT LOG · 4 RECORDERS · 11 DAYS
One-minute recordings logged12,843
Gaps in recording0
Species in 30 minutes, ours vs theirs11 vs 2
Species confirmed across the site35
Impossible species removed26 of 66
Detections made at night35%
Likely faulty recorder, fewer detections13.5×
Recorder pops repaired in the audio sample99.8%
Possible-speech minutes in the resultsNONE
THE PARTNER AND SITE ARE NOT NAMED. RESULTS ARE SHARED WITH THE PARTNER'S AGREEMENT.
[ 05 · FIELD TEST ]
As little as six hours can profile a site.
On 18 October 2026 we recorded from 07:31 to 13:46 at a site near Beri, Rajasthan, about 30 km west of Keoladeo National Park. The signals in the header above are from this recording.
Those six hours revealed 13 bird species, Osprey among them, and no threats. They also showed an engine running throughout and a highway 100 m away masking quieter calls, which is why the next recorder there will go well away from the road.
From the same audio we estimate a recorder hears a chainsaw out to about 110 m and a small songbird to about 40 m. Measured per site, ranges like these set how far apart recorders go.
FIELD LOG · BERI · 18.10.26
SPECIESFIRST HEARDBEST CONF
Red-wattled Lapwing07:320.99
Rose-ringed Parakeet07:410.98
Rufous Treepie07:440.91
Common Chiffchaff08:260.90
Osprey07:500.89
White Wagtail08:010.89
White-throated Kingfisher07:410.87
Common Tailorbird07:560.83
+ 5 more species
Engine hum, 45 to 80 HzthroughoutBASELINE
Threats detectedNONE
CONFIDENCE IS THE MODEL'S SCORE, NOT A HUMAN CHECK. RANGES ARE ESTIMATES FROM THIS RECORDING, NOT A GUARANTEE FOR EVERY SITE.
Most systems classify what they know. Canopy also listens for what it does not.
Every site gets its own baseline, learned from its own background sound. A sound does not need a name to be worth a second listen.
[ 06 · WORKING WITH PARTNERS ]
Built for land that people share.
01 / OWNERSHIP
Your recordings stay yours
Audio from your site belongs to you. We use it for your reports, and to improve our models only where our agreement allows it.
02 / PRIVACY
Conversations stay private
A speech check flags minutes that may contain people talking. Nobody listens to them, they are never shared, and they are left out of partner reports. We listen for wildlife and threats, not conversations.
03 / FIRST LOOK
You hear it first
Results go to the site partner before anyone else. Nothing about your site is published without your agreement, and nothing that names you or your site without your written consent.
[ 07 · PROGRESS ]
Where Canopy is today.
An honest timeline rather than a polished promise.
-
DONE
Bird model
A bird model that outscores BirdNET 3.0 on held-out field recorders.
-
DONE
First field test
Six hours at Beri, Rajasthan, analysed end to end.
-
DONE
First partner deployment
A signed MOU with an Indian restoration nonprofit, and a review of their four-recorder, eleven-day lake deployment, reported back to them.
-
IN PROGRESS
The Canopy recorder
Our own recorder board is designed and out for fabrication quotes. Two pilot nodes built on an off-the-shelf board are on the bench.
-
IN TESTING
Threat detector
A per-site threat detector with a review list for flagged events, being tested on real field audio.
-
NEXT
Multi-site deployments
Recorders across partner sites, with repeat deployments to measure change over time.
[ 08 · WHO IT IS FOR ]
For the people responsible for the land.
01 / RESTORATION
Restoration and CSR programmes
Evidence of which species use the sites you fund, and with repeat deployments whether that is changing, not just a sapling count.
02 / PROTECTION
Forest departments and reserves
A record of activity at the edges of protected land, with unusual sounds flagged for review.
03 / RESEARCH
Ecologists and researchers
Species detections with confidence scores from long, unattended recordings at your study sites.
[ 09 · TEAM ]
A small team building Canopy, with field work in India.
FOUNDER
Arjun Tyagi
HARDWARE
Hriday Matharu
FINANCE
Siyona Sara
MARKETING
Tashi Shrivastava
[ 10 · OPEN A CHANNEL ]
Have land worth listening to? Tell us about it.
Where it is, roughly how big, and what you want to know about it. Research collaborations and press use the same address.