PatiPix/AudioSeal-ONNX

AudioSeal generator (16 bits) — ONNX

ONNX export of the AudioSeal watermark generator audioseal_wm_16bits (facebook/audioseal, revision 3c19eba53390776cf2cc9ed5f6c9ac67ce72ecba, file generator_base.pth), made with the official AudioSeal 0.2.0 code.

PatiPix Studio adds this invisible watermark to every voice-over it creates, to mark it as AI-generated (EU AI Act, article 50), together with a signed C2PA manifest.

Inputs and outputs

Name	Shape	Type	Meaning
son (input)	[1, 1, T]	float32	audio at 16 kHz; T must be a multiple of 320
message (input)	[1, 16]	int64	the 16-bit message (0 or 1)
filigrane (output)	[1, 1, T]	float32	the watermark, to add to the audio

How it was made

torch 2.10.0, TorchScript export (dynamo=False), opset 17, weights unchanged. AudioSeal's small padding steps (useless when T is a multiple of 320) were replaced by their value (zero) for the export.

Checked on 12 voice clips: difference between ONNX and PyTorch below 2e-6; the official detector finds the watermark on 100 % of the audio and reads the message back.

File

File	Size (bytes)	SHA-256
audioseal_generator_16bits.onnx	58,814,395	622ce07fdac2cf2faa83be25ba2ed3f830da27026746755a66859bd8ba2410ed

License and credits

MIT License, like AudioSeal (see LICENSE). Copyright (c) Meta Platforms, Inc. and affiliates. AudioSeal: San Roman et al., « Proactive Detection of Voice Cloning with Localized Watermarking », ICML 2024.