File size: 1,752 Bytes
80b60f9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 | ---
license: mit
tags:
- watermarking
- audio
- safetensors
library_name: safetensors
---
# perth-implicit
A converted **safetensors** mirror of Resemble AI's **PerTh** implicit neural watermarker
(`PerthImplicitWatermarker` / PerthNet — the `perth_net_250000.pth.tar` "implicit" run) from the
MIT-licensed [`resemble-perth`](https://github.com/resemble-ai/Perth) package, version **1.0.1**.
This repository is a redistribution mirror published by **SceneWorks** so the weights can be resolved
by an immutable commit SHA. It exists to serve SceneWorks Inference's native-candle PerTh watermarker
(`candle-audio-chatterbox`, `crate::perth`), which applies provenance watermarking to Chatterbox
voice-clone output. It is **not** an official Resemble AI distribution.
## Contents
- `perth_implicit.safetensors` — the encoder + decoder convolution weights, 56 tensors, `float32`.
The original checkpoint's `ap.*` STFT window buffers are **omitted** because they are recomputable
constants (the consuming Rust port recomputes the Hann window).
- `sha256`: `5fc2f30f327b1ba4b3ed760b9e0e46afff7aa6eb73681b1b8a8154d7768346db`
- `LICENSE` — the upstream MIT license, preserved verbatim.
## Provenance
- **Source**: `resemble-perth` 1.0.1 → `perth/perth_net/pretrained/implicit/perth_net_250000.pth.tar`
- **Conversion**: `scripts/audio/convert_perth_watermarker.py` (a torch-free converter) in the
SceneWorks Inference repository. The conversion copies the raw float32 conv storages verbatim.
- **Upstream project**: https://github.com/resemble-ai/Perth
## License
**MIT** — Copyright (c) 2025 Resemble AI. See `LICENSE`. This mirror preserves the upstream license
unchanged; all rights and attribution remain with Resemble AI.
|