Problem
Passive acoustic monitoring networks generate millions of hours of recordings, and audio captures the nocturnal and cryptic species that cameras and satellites miss. Species classifiers (BirdNET, Perch) are mature. Turning “the soundscape shifted in March” into “here’s what likely happened and what to check next” is still manual, expert-time-intensive work.
System
Ingest → trend → attribute → assess → write. Every index series is deseasonalised against a robustly fitted harmonic model of day-of-year and day-of-week before change-point detection, because on a raw tropical series a detector finds the annual cycle and nothing else. Candidate explanations are retrieved from the site’s field log and the ecological literature. A rubric owns the verdict and the confidence; the language model ranks hypotheses and writes prose, and never emits a score. A review dashboard shows the ledger.
Worth knowing
- Confounds compete as hypotheses rather than being filtered. The biggest shift in most deployments is an equipment change. A gain change moves level-dependent indices while ratio-based ones hold steady, and that split is diagnosed explicitly.
- A claim linter gates publication. Every sentence must cite evidence that exists, every number must appear in a cited payload, and language strength must match the confidence tier. Violations are a reported metric, because overclaiming is the failure mode that destroys trust in a tool like this.
- A model asked to judge and report its own certainty produces a number nothing can falsify. Splitting those roles is the central design choice.