zerds-public

Public assets that Zerds downloads at runtime. This repository exists only because the main Zerds repository is private: GitHub release assets inherit their repository's visibility, so an app running on a user's machine — an anonymous client — cannot fetch them.

Nothing proprietary lives here. Everything in this repository is either an open-source model redistributed under its own license, or a file Zerds needs to be publicly reachable.

Contents

Path	What it is	License
models/yamnet/model.onnx	YAMNet audio-event classifier, converted from the official TensorFlow Hub model to ONNX. Takes a 1-D 16 kHz mono waveform and returns 521 AudioSet class scores per ~0.48 s frame. Zerds uses it to mark laughter, applause, cheering, crying, coughs and sneezes on the timeline.	Apache-2.0 — see models/yamnet/LICENSE and models/yamnet/NOTICE

Integrity

Every file here is pinned by SHA-256 in the Zerds source. A download whose digest doesn't match the pinned value is rejected before it reaches the runtime, so replacing a file in this repository does not silently change what users run — it makes their download fail until Zerds is updated to match.

Current digests:

9a6aef46391df10a4ff804ea196266855a532a200ec13894024da24e16aaf7b1  models/yamnet/model.onnx

Verify locally with:

shasum -a 256 models/yamnet/model.onnx

Adding files

Two constraints worth knowing before adding anything large:

• GitHub rejects files over 100 MB (and warns above 50 MB).
• Git keeps every version forever. Replacing a binary does not free the old one; the repository only grows, and every clone pays for it.

So this tree suits assets in the tens of megabytes. For anything larger — the Whisper models are hundreds of MB, GGUF language models are gigabytes — use a GitHub release on this repository instead: release assets are public here too, are served from a CDN, have no practical size limit, and stay out of git history.

How the YAMNet model was produced

Not hand-converted. .github/workflows/build-yamnet-onnx.yml in the Zerds repository converts the official Apache-2.0 TF-Hub model with tf2onnx and validates the result in the same run — graph shape, frame counts against YAMNet's 0.96 s window / 0.48 s hop, and numeric parity against the original TensorFlow model. It refuses to publish anything that fails. The build that produced this file reported a maximum score difference of 1.2e-06 against TensorFlow and 100% top-class agreement across four waveform lengths.