Talk2Nature
← Research libraryField notes

Machine learning / Software documentation

Perch: reusable representations of animal sound

Source year 2025 · Note reviewed 2026-10-14

Perch v2 provides acoustic embeddings for downstream classifiers. The official CPU version 1 card declares Apache-2.0 and says its weights match the GPU perch_v2 version 2. Pinned Hoplite configuration uses five-second, mono 32 kHz windows and a 1,536-dimensional embedding.

What we reviewed · Documentation checked

Official parent card and exact perch_v2_cpu/1 variation page reviewed, along with the repository migration notice, Hoplite README and pinned model configuration. No weights downloaded and no inference run.

Not yet independently reviewed by a scientific specialist.

What this does — and doesn’t — establish

Google's card names Xeno-Canto, iNaturalist, Animal Sound Archive and FSD50k as training sources, but the item-level training data and license provenance, target-data overlap and scientific fit remain unverified. The card cautions that logits are uncalibrated and may be unreliable for rare species. No bundle checksum, runtime, accuracy or mobile performance was verified.

What we can build on

Consider a frozen embedding baseline once the species, labels and data rights are established.

See the related tools and reading paths →

The next reading task

Before selecting this baseline, verify the exact package integrity and applicable download terms, inspect preprocessing, then measure a CPU inference path on synthetic or rights-cleared audio. Keep behavior/context benchmark admission gated on independently reviewed labels and leakage-safe splits.

Go to the sources

  1. Perch 2.0 ↗

    Google DeepMind. How AI is helping advance the science of bioacoustics to save endangered species (7 August 2025). Reusable bioacoustic representations.

  2. Perch source code ↗

    Google Research. Perch repository, accessed 30 September 2026. Code and links to model artifacts; inspect the exact weight license before deployment.

  3. Perch 2 model card ↗

    Google. Perch 2 TensorFlow model card: release variants, input/output, limitations and Apache-2.0 terms. No weights downloaded.

  4. Perch-Hoplite tooling ↗

    Google Research. Official inference and agile modeling repository, reviewed at commit 050c57b5dac2f6d28ab1f562fa2000c8dcbc2302.

  5. Perch 2 CPU variation v1 model card ↗

    Google. Exact Perch v2 CPU variation and version page. Declares Apache-2.0, CPU compatibility, and matching weights to the GPU perch_v2 version 2. No weights downloaded.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.