cstr/audioseal-GGUF

AudioSeal (GGUF)

GGUF conversion of AudioSeal (audioseal_wm_16bits generator + audioseal_detector_16bits detector) for use as the optional neural audio-watermarking backend in CrispASR.

AudioSeal is a proactive, localized watermarking method for AI-generated speech: an imperceptible watermark is embedded at synthesis time and recovered by a fast detector, robust to common audio edits (re-encoding, compression, resampling).

Contents

File	Contents	Size
audioseal.gguf	Generator (73 tensors) + detector (40 tensors), F32	~89 MB

Sample rate 16 kHz, 16-bit message payload.

Usage (CrispASR)

# Opt in to the neural watermark (otherwise CrispASR uses its built-in
# zero-dependency spread-spectrum watermark):
crispasr --backend <tts-backend> -m auto --tts "..." \
    --watermark-model auto            # auto-downloads this GGUF

# Detect:
crispasr --detect-watermark out.wav --watermark-model auto

License & attribution

This GGUF is a format conversion of AudioSeal, released by Meta AI (FAIR) under the MIT License — both code and model weights. See the upstream repository for the original weights and paper:

• Code/weights: https://github.com/facebookresearch/audioseal (MIT)
• Paper: Proactive Detection of Voice Cloning with Localized Watermarking, Roman et al., ICML 2024 — https://arxiv.org/abs/2401.17264

Redistributed here under the MIT License with attribution to the original authors. No model weights were retrained or modified; only the container format was changed (PyTorch → GGUF).