Hugging Face's logo

SceneWorks
/
perth-implicit

Safetensors
watermarking
audio
Model card Files
xet
 Community
perth-implicit / README.md
SceneWorks's picture
SceneWorks
Add PerTh implicit watermarker (converted from MIT resemble-perth 1.0.1) + LICENSE + provenance README (sc-13443)
80b60f9 verified
preview code
|
Raw
History Blame Contribute Delete
1.75 kB
metadata 
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 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.