[ 01 · THE PIPELINE ]
Eight steps, one recording at a time.
-
01
IN THE FIELD
Record
A recorder saves one minute of sound every five minutes to a memory card, day and night. It only records. No analysis runs on the device.
ONE MINUTE EVERY 5 MIN
288 MINUTES A DAY
-
02
IN THE FIELD
Collect
When the cards come back, every file is logged: which recorder, what time, battery level, and whether any minute is missing. Gaps and power problems show up here, before any analysis.
PARTNER SITE
12,843 MINUTES, NONE MISSING
-
03
IN THE CLOUD
Speech check
Every minute of audio we receive is checked for people talking by a sound model. Minutes that may hold speech are flagged: nobody listens to them, they are never shared, and they are left out of the partner's results. A gate that drops these minutes before any other model runs is planned.
RUNS TODAY
GATE PLANNED
A SIMPLE VERSION RUNS IN THE DEMO
-
04
IN THE CLOUD
Repair
Recorder glitches, such as the pops an overloaded recorder writes when its samples wrap around, can be repaired so they are not mistaken for real sounds such as gunshots.
TESTED ON A PARTNER SAMPLE
NOT YET IN THE PIPELINE
99.8% OF POPS REMOVED
-
05
IN THE CLOUD
Species
Our bird model names the species in each window and gives each a confidence. Species with no recorded occurrence in the region are removed. At the partner site, that check took out 26 of the 66 species an off-the-shelf pipeline had reported.
+20% VS BIRDNET 3.0
WITH SITE TRAINING · BENCHMARKS
-
06
IN THE CLOUD
Threats
We are testing a threat check in which our sound model scores every window for chainsaws and gunshots. The alert line is set from each site's own background and never drops below a floor, so an engine that runs all day does not set it off. Fire is named but goes to review, not to an alert, until it has been tested at a site.
IN TESTING
BERI REPLAY
~1.6 FALSE ALERTS A DAY
A SIMPLE VERSION RUNS IN THE DEMO
-
07
PEOPLE
Review
Sounds that are out of place for the site, anything named fire, and would-be alerts that sound more like rain or thunder go to a review list, with the audio, for a person to hear. Every alert comes with its audio too.
IN TESTING
BERI REPLAY
~4.8 REVIEW ITEMS A DAY
-
08
PEOPLE
Report
A plain report: which species were heard and how sure we are, day versus night, every alert and review item with its audio, and the health of each recorder. It goes to the site partner before anyone else.
TO THE PARTNER FIRST
[ 02 · TRY THE MODEL ]
Run our sound model on your own clip.
This is the sound model behind step 06, and it also runs a simple version of the step 03 speech check. It uses the same weights we run on field audio, partly compressed to load in a browser. Drop in a recording, or record ten seconds where you are.
It names general sounds, such as birds, insects, rain or vehicles, rather than bird species. On our Beri clips it hears vehicle and engine noise in most windows: an engine ran near the recorder throughout, and a highway is about 100 m away. That is why the next recorder there will go well away from the road.
It runs in your browser. Your clip is never uploaded or saved. Only the model files (92 MB) stay in your browser's storage, so later visits skip the download.
LIMITS OF THIS DEMO
WITH NO SITE BACKGROUND, THE DEMO USES OUR LOWEST ALERT LINE, SO IT LEANS TOWARD ALERTING. IN THE FIELD VERSION, NOW IN TESTING, THE LINE IS SET FROM EACH SITE'S OWN SOUND, AND A SECOND CHAINSAW SCORE IS ADDED THAT CATCHES MORE DISTANT SAWS. SPECIES COME FROM OUR SEPARATE BIRD MODEL AND APPEAR IN REPORTS. BEST ON A LAPTOP OR DESKTOP.
OR A CLIP FROM OUR BERI FIELD TEST
Choose a clip or record one. The first run downloads the model (92 MB, once), then each 5-second window takes a second or two.
SOUND MODEL: PANNS CNN14 (KONG ET AL., 2020), CC BY 4.0, EXPORTED TO ONNX AND PARTLY COMPRESSED TO 8-BIT BY CANOPY. SOUND NAMES: GOOGLE AUDIOSET ONTOLOGY, CC BY-SA 4.0. RUNTIME: ONNX RUNTIME WEB, MIT. FULL CREDITS. PEOPLE'S SOUNDS ARE NEVER SHOWN. PERCENTAGES ARE THE MODEL'S OWN SCORES, NOT A HUMAN CHECK.
[ 03 · OPEN A CHANNEL ]
Want this on your land? Tell us about it.
Where it is, roughly how big, and what you want to know about it.