thapaee/audioseal-onnx

AudioSeal (16-bit) in ONNX

Meta's AudioSeal audio watermark, the 16-bit generator and detector, converted to ONNX. VoicedByMe uses it to mark the speech it generates as AI-made. The weights are Meta's, unchanged: only the file format is different. Code and weights are under the MIT license (see LICENSE).

File	What it does	SHA-256
audioseal-16bit-generator.onnx	Makes the watermark for a clip	c7954ebc5a097c41c10b099d80e1cb218168d79ed56f19138b1f43d705a1524b
audioseal-16bit-detector.onnx	Finds the watermark and reads its message	02856c2d372472314c674fb000a1844c6c38b0a75703371dd9255f83e2a841c1

How it was made

From the audioseal 0.2.0 package's checkpoints (audioseal_wm_16bits, audioseal_detector_16bits, from facebook/audioseal), exported with PyTorch's ONNX exporter (opset 17). The model's padding is worked out for lengths that are a multiple of 320 samples, so the graph works for any such length. On those lengths the results match PyTorch's to within about 1e-7.

Using it

Both models take 16 kHz mono audio, (1, 1, T) float32, where T is a multiple of 320 (pad with zeros and cut back afterwards).

• Generator: inputs audio and message (1, 16) int64 (the bits). Output watermark (1, 1, T); the marked audio is audio + watermark.
• Detector: input audio. Outputs prob (1, T) (how likely each sample is marked) and bits (1, 16, T) (bit logits; average over time, then > 0 means 1). The message reads back only when the audio is on the same 20 ms grid as when it was marked, so for audio that may have been shifted (some encoders add a delay), try offsets of 0 to 319 samples.