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.

Files

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	MatMul-only dynamic INT8 (arm / low-RAM). ~64% less peak RAM than fp32 (292 MB vs 816 MB on Pi5), ~34% faster, identical top-1 and top-5 vs fp32. Has the DFT front-end, so ONNX Runtime only.
`labels.txt`	313 KB	Species labels (14,795 entries, iNaturalist taxonomy)
`SHA256SUMS`	-	Checksums for integrity verification

Model Information

Input

Name: inputs
Shape: [batch, 160000] (5 seconds at 32 kHz)
Type: float32

Outputs

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

Species Coverage

Approximately 10,000 bird species
Frogs, crickets, grasshoppers, and mammals
Based on training data from Xeno-Canto, iNaturalist, Animal Sound Archive, and FSD50k

Performance (ONNX vs TFLite, 100 runs)

Metric	ONNX	TFLite
Mean	66.4 ms	608.8 ms
Speedup	9.2x faster	baseline

License

Apache 2.0, following the original Google Perch license.

Credits

Google Research for the original Perch v2 model
justinchuby for the ONNX conversion and DFT-to-MatMul optimization
cgeorgiaw for the Keras-compatible model and label files

References

Perch model: Google Bird Vocalization Classifier on Kaggle
ONNX source: justinchuby/Perch-onnx
Labels and Keras model: cgeorgiaw/Perch
Perch codebase: google-research/perch

tphakala/Perch-v2

Google Perch v2 (ONNX)

Origin and Attribution