Inference Engine

ONNX Runtime integration and model inference.

Architecture

InferenceIsolate (background Dart isolate)
  └─ InferenceService (high-level coordinator)
      └─ ClassifierModel (low-level ONNX session wrapper)

ClassifierModel

Direct ONNX Runtime FFI wrapper:

• loadModel(bytes) / loadModelFromFile(path) — create OrtSession
• predict(audioSamples, windowSamples) — run inference, return ModelOutput
• Resolves output tensor indices by name (handles any graph ordering)

InferenceService

High-level coordinator:

• initialize(modelBytes, labelsCsv, config) — load model + parse labels
• infer(audioSamples) — model → temporal pooling → sensitivity → top-K
• Log-Mean-Exp temporal smoothing across recent windows (α=5.0, max 5 windows)

InferenceIsolate

Runs InferenceService in a background Dart isolate:

• start(modelFilePath, labelsCsv, config) — spawn isolate, load model
• infer(audioSamples) — send request via SendPort, await response
• stop() — kill isolate and clean up

The model file path (not bytes) is sent to the isolate to avoid serializing ~259 MB.

Geo-Model

GeoModel runs a smaller ONNX model (~7 MB) on the main thread:

• Input: [latitude, longitude, week] (normalized)
• Output: per-species probability scores
• Used for geographic species filtering

Model Config

All model parameters are in assets/models/model_config.json:

• Tensor names, sample rate, window duration
• Label file format (delimiter, columns)
• Inference defaults (sensitivity, threshold, top-K, pooling)