synath/audioseal-onnx

AudioSeal 16-bit — ONNX export

ONNX exports of Meta's AudioSeal 16-bit generator and detector, for inference with onnxruntime (Node / web / Python) without a PyTorch dependency.

File	Role
audioseal_generator_16bits.onnx	Embeds a 16-bit watermark into mono audio.
audioseal_detector_16bits.onnx	Detects the watermark and recovers the 16-bit payload.

Usage notes (important)

• Inputs are float32 mono PCM in [-1, 1]. Pad the audio length up to a multiple of HOP = 320 before running each model, then trim the output back to the original length.
• Generator: inputs audio [1, 1, T] (float32) + message [1, 16] (int32 bits); output is the watermarked audio.
• Detector: input audio [1, 1, T] (float32); outputs per-sample detection probabilities + the recovered 16-bit message.
• Sample-rate-agnostic and length-dynamic — one generator + one detector file covers any sample rate and clip length.

Provenance

Exported from the audioseal PyPI package's PyTorch checkpoints via torch.onnx (opset 17), with NO_TORCH_COMPILE=1 and a host-side pad-to-320 / trim convention. Bit-exact 16-bit payload parity and matching detection probabilities versus the PyTorch reference were verified.

License

MIT — the same license as upstream AudioSeal (code and model weights, which Meta relicensed to MIT in April 2024). These ONNX files are a derivative of those MIT-licensed weights, redistributed under MIT. © Meta Platforms, Inc. and affiliates. See LICENSE.