iNaturalist Sound Classifier 作成者: biodiversica

What does it do?

When you visit an iNaturalist observation page that contains a sound recording, this extension adds a panel that lets you run AI-powered bioacoustic analysis with a single click. It helps identify which species are vocalizing, validates those detections against geographic occurrence data, and displays ranked results with confidence scores. Whether you're a researcher or a citizen scientist, this tool can help improve sound-only data identification on iNaturalist.



Key Features

->Runs 100% locally
All AI inference happens in your browser using WebAssembly.
There are no accounts, no subscriptions, and no data collection.

->Geographic check
The extension reads the observation's coordinates and automatically filters available models by region. After generating predictions, it cross-references every top detection against GBIF and iNaturalist species occurrence databases to determine whether that species has been documented in that area.
Detections are marked as:
- ✓ — within known range
- ⚠ — outside known range
This helps distinguish likely identifications from unusual records.

->State-of-the-art models
Comes pre-configured with two leading bioacoustic models:
- BirdNET v2.4 (Cornell Lab of Ornithology) Trained on over 6,000 bird species worldwide; one of the most widely used bird sound classifiers.
- Perch v2.0 (Google Research) A broader-scope model covering animal vocalizations across taxonomic groups. Both models are downloaded on first use and cached locally, so subsequent analyses load instantly.

->Fully configurable
- Adjust the confidence threshold to filter weak detections
- Control analysis window overlap for finer time resolution
- Choose between softmax, sigmoid, or raw logit outputs
- Export results as CSV for downstream analysis
- Clear the model cache at any time from the settings panel

->Extensible: bring your own models
Researchers and developers can add any ONNX classification model through the extension's model manager UI — no code changes required.
Configure:
- Model URL (huggingface or zenodo)
- Sample rate
- Window size
- Label file
- Activation function
The extension handles downloading, caching, and inference automatically.



How it works

1. Navigate to any iNaturalist observation page that has a sound recording attached.
2. The extension panel appears automatically. Select a model from the list (filtered by location relevance).
3. Click Run Analysis. The model downloads on first use, then analysis begins immediately.
4. Audio is fetched from iNaturalist, decoded, resampled to the model’s required sample rate, and split into overlapping time windows.
5. Each window is processed by the AI model running locally via ONNX Runtime WebAssembly.
6. Top predictions are validated against geographic occurrence data from GBIF and iNaturalist.
7. Results are displayed in ranked order with species names, confidence scores, time windows, and range validation status.
8. Optionally export results as CSV for further analysis.

Model licenses

• BirdNET v2.4: CC BY-NC-SA 4.0
  (Cornell Lab of Ornithology / Chemnitz University of Technology)
• Perch v2.0: Apache 2.0
  (Google Research)

This extension is open source (GPL-3.0). Contributions and custom model configurations are welcome.



Privacy

This extension does not collect, transmit, or store any personal data.
- Audio is fetched directly from iNaturalist’s public API (the same request your browser makes when you press play)
- Processing happens entirely locally in your browser
- The only outbound requests are to iNaturalist and GBIF public APIs for species occurrence metadata
- No audio data is transmitted externally
No analytics. No tracking. No third-party services.