ONNX format of the Google Perch v2 bird vocalization classifier, repackaged for use with BirdNET-Go.
The Perch v2 model was developed by Google Research as part of the bird-vocalization-classifier project. It uses an EfficientNet-B3 architecture with approximately 12 million embedding parameters and 91 million classification parameters covering nearly 15,000 species.
The ONNX conversion was performed by justinchuby, who also created an optimized variant with the DFT node converted to MatMul for additional speedup.
The label file originates from the cgeorgiaw/Perch repository on HuggingFace.
| File | Size | Description |
|---|---|---|
perch_v2.onnx |
409 MB | Standard ONNX model (tolerance 1e-5 vs TFLite) |
perch_v2_no_dft.onnx |
413 MB | ONNX with DFT converted to MatMul, faster inference (tolerance 2e-4) |
perch_v2_int8_arm.onnx |
131 MB | Withdrawn 2026-11-15, kept for existing installs only. MatMul-only dynamic INT8: against fp32, top-1 agrees on 82% of 5,560 real station windows, below the publish gate's int8 floor. Use perch_v2_no_dft.onnx. |
labels.txt |
313 KB | Species labels (14,795 entries, iNaturalist taxonomy) |
SHA256SUMS |
- | Checksums for integrity verification |
inputs[batch, 160000] (5 seconds at 32 kHz)float32| Output | Shape | Type | Description |
|---|---|---|---|
embedding |
[batch, 1536] |
float32 | Audio embedding vector |
spatial_embedding |
[batch, 16, 4, 1536] |
float32 | Spatial features |
spectrogram |
[batch, 500, 128] |
float32 | Computed spectrogram |
label |
[batch, 14795] |
float32 | Species classification logits |
| Metric | ONNX | TFLite |
|---|---|---|
| Mean | 66.4 ms | 608.8 ms |
| Speedup | 9.2x faster | baseline |
Apache 2.0, following the original Google Perch license.
perch-v2-fp32-no-dft.onnx 413 MBperch_v2_no_dft.onnx 413 MBperch-v2-fp32.onnx 409 MBperch_v2.onnx 409 MBperch-v2-int8-arm.onnx 131 MBperch_v2_int8_arm.onnx 131 MBcgeorgiaw/Perch