Syrra
AI for acoustic biodiversity inventories.
Result
- 1st prize, D4Gen 2026 — Genopole, 48 h, €10,000
- First contract signed two months later, with a global cloud provider (data-center permitting in France)
- Four evaluations under way with CAC 40 groups
The problem
- Every large project in France needs an authorization backed by a defensible fauna inventory (~5,000 filings a year)
- One hour of audio takes thirty minutes of expert listening; one site produces ~23,000 hours a year
- Nobody listens to it all — they sample, which is a legal weakness
The product
- Microphones record a site continuously; the pipeline processes 100% of the recordings
- A fine-tuned, state-of-the-art detection model that outperforms BirdNET
- Detections are clustered; the ecologist validates clusters instead of listening to hours of audio
- 100 h of audio → minutes of review; every detection stays auditable
