Assa call recogniser: Perch 2 + random forest (V1)
Detects calls of Assa in field recordings. Each one-second chunk of audio is turned into a Perch 2 embedding (Google DeepMind's model for animal sounds), and a random forest scores the embedding for the target call. Chunks scoring above 0.375 count as detections.
A fine-tuned Whisper recogniser for the same species is at LBolitho/Assa_Call_Recogniser_V1. The Multi-Species Call Recogniser notebook can run either model, or both.
Files
| File | What it is |
|---|---|
perch_rf.joblib |
the trained random forest (scikit-learn) |
perch_rf_info.json |
the settings the model must be used with |
How the audio must be prepared
- Convert to mono at 32 kHz.
- Cut into 1000 ms clips; centre each clip in a zero-padded 5 s window.
- Embed with Perch 2 (
perch_hoplite.zoo.model_configs.load_model_by_name("perch_v2")) and average the embedding over frames. - Score with
rf.predict_proba(embeddings)[:, 1].
Training
- Random forest: 500 trees
- Classes: {'0': 'Non_Target_sounds', '1': 'Target_sounds'}
- scikit-learn 1.6.1 (load the model with this version)
Validation
On held-out one-second clips from the training recordings:
| threshold | precision | recall | f1 | npv | roc_auc | avg_precision | |--------------------- | 0.375 | 1.000 | 0.994 | 0.997 | 0.999 | 1.000 | 1.000 | | 0.500 | 1.000 | 0.987 | 0.994 | 0.997 | 1.000 | 1.000 |
With whole recordings held out (grouped 5-fold cross-validation): F1 0.998, average precision 1.000.
Limitations
Tested only in the habitats where the training recordings were made. Validation clips come from the same recordings as the training clips, so field performance should be checked with manual review.
perch_rf.joblib is a pickle file: only load it from this repository.
- Downloads last month
- -