Arm/perch-v2-int8-litert

Perch V2 8-bit

Perch is a bioacoustics model trained to classify nearly 15,000 species, of which ~10,000 are birds.

✨ Key Highlights

• Optimized for edge deployment — A quantized 8-bit variant of the Perch model capable of running in real time on Raspberry Pi 5.

📦 Model Details

Model Description

This package contains a modified variant of the Perch V2 audio classifier distributed in the ernensbjorn/perch-v2-int8-tflite repository. It accepts a batch containing one five-second, 32 kHz audio clip and returns probabilities for 14,795 classes together with intermediate model representations.

• Developed by: Google (upstream model)
• Model type: Audio classifier
• License: Apache-2.0
• Base model: ernensbjorn/perch-v2-int8-tflite
• Packaged variant: 8-bit LiteRT model with softmax output

Model Sources

• Base model artifact: perch_v2_original.tflite at revision 71fd62f8144c6214e494da9a70065b2ce55a5ad4
• Upstream Perch V2 model: Google Bird Vocalization Classifier

🚀 Get Started with the Model

🔓 Compute Flow — Early Access

The inference engine for this model package is available through the Compute Flow Early Access Program.

Want to try it?

📩 Contact us at ai-early-access@arm.com to request access.

📊 Quality evaluation

For accuracy results, refer to the ernensbjorn/perch-v2-int8-tflite model card.

🎯 Performance evaluation

Performance results for this package will be reported in a future update.

🛠️ Technical Specifications

Objective

The package provides Perch V2 for audio detection and classification.

Runtime Architecture

Component role	Framework / format
Audio classifier	LiteRT (.tflite)
Runtime configuration	Audio-detection manifest (.json)
Class-name mapping	JSON
Representative input	Mono, 32 kHz, float32 WAV

Precision and Quantization

The packaged artifact is the repository's 8-bit LiteRT variant. Its public tensor interface uses float32 for the audio input and all four outputs.

Input Specification

Input	Description
audio	Float32 tensor with shape [1, 160000], representing one five-second audio clip sampled at 32 kHz

Output Specification

Output	Description
spectrogram	Float32 tensor with shape [1, 500, 128]
spatial_embedding	Float32 tensor with shape [1, 16, 4, 1536]
embedding	Float32 tensor with shape [1, 1536]
label	Float32 tensor with shape [1, 14795] containing class probabilities

Repository Contents

• assets/class_names.json — class-name mapping referenced by the manifest.
• perch_v2_8bit.tflite — LiteRT audio-classification model referenced by the manifest.
• perch_v2_8bit_manifest.json — audio-classifier manifest.
• samples/erithacus_rubecula.wav — audio sample for testing the model with audio_detection_cli.
• SHA256SUMS — model-package checksums for reproducibility.

🗂️ Model and Asset Origin

• Model: perch_v2_8bit.tflite was derived from perch_v2_original.tflite. The following changes were made for this package:

  • Set the batch dimension to 1.
  • Added a softmax layer to the label output so it returns probabilities.
  • Corrected the label signature entry to reference the actual subgraph output tensor.
  • Replaced 365 explicit BROADCAST_TO operations with equivalent implicit broadcasting and removed unreachable shape calculations.

• Class names: assets/class_names.json was transformed from labels.txt.

• Sample: samples/erithacus_rubecula.wav was created from Wikimedia Commons' European-robin recording, credited to Gilles San Martin, with:

  ffmpeg -y \
      -i 120401-132827_Erithacus_rubecula.ogg \
      -ac 1 \
      -ar 32000 \
      -t 55 \
      -af apad \
      -c:a pcm_f32le \
      samples/erithacus_rubecula.wav
  

🔐 Checksums

SHA256SUMS was generated by recursively hashing every regular file in the model bundle, including files in subdirectories, except the generated root SHA256SUMS and paths with a dotfile component.

From the model bundle root, verify the checked-out files with:

shasum -a 256 -c SHA256SUMS

Citation

If you use this model, cite the original Perch work:

@article{ghani2023global,
  title={Global birdsong embeddings enable superior transfer learning for bioacoustic classification},
  author={Ghani, Burooj and Denton, Tom and Kahl, Stefan and Klinck, Holger},
  journal={Scientific Reports},
  year={2023}
}