High-performance acoustic embedding extraction and species classification using Google's Perch v2 model for large-scale bioacoustics inference on CPUs.
- High-performance CPU inference — Up to 750x realtime with ONNX or TFLite
- Batch processing — Large-scale dataset processing with checkpointing and resume
- Visualization — Spectrogram overlays and detection time series charts
# Install dependencies
pip install -r requirements.txt
# Download model
wget https://huggingface.co/justinchuby/Perch-onnx/resolve/main/perch_v2.onnx -P models/perch_v2/
# Run inference
python scripts/inference/perch-onnx-inference.py --audio-dir ./data/test-data --output-dir ./output| Guide | Description |
|---|---|
| Usage Guide | Installation, running inference, visualization, output formats, benchmarks |
| Batch Processing | Large-scale inference, monitoring, log parsing |
- Perch v2: Google's classifier pre-trained on a multi-taxa dataset (Paper)
- Input: 5-second audio chunks at 32kHz
- Output: Species predictions and 1536-dimensional embeddings
| Format | Source | Prepared by |
|---|---|---|
| ONNX | Hugging Face | Justin Chu (@justinchuby) |
| TFLite | Bioacoustics Model Zoo | Lapp, S., and Kitzes, J. (2025) |
├── scripts/inference/ # Inference scripts
│ ├── perch-onnx-inference.py # Main ONNX inference script
│ └── perch-tflite-inference.py # TFLite inference script
├── scripts/ # Benchmarking and visualization tools
├── tools/ # Shell scripts for batch processing
├── docs/ # Documentation
├── models/ # Model files (downloaded separately)
└── data/ # Test data
See LICENSE file for details.
Developed as part of the PhenoScale project, UKRI Frontiers grant EP/X024520/1 awarded to Ben Sheldon, University of Oxford.