Implementation:Facebookresearch Audiocraft AudioSeal WMModel

Knowledge Sources	Facebookresearch_Audiocraft AudioSeal
Domains	Audio_Watermarking, Model_Architecture
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for audio watermark embedding and detection using the AudioSeal architecture.

Description

This module defines the abstract WMModel interface for audio watermarking and provides the AudioSeal implementation. WMModel defines the API with get_watermark, detect_watermark, and nbits property. AudioSeal wraps the audioseal library's generator and detector models, handling message encoding and detection output parsing.

Usage

Import this class when you need to embed or detect audio watermarks using a trained AudioSeal model.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/models/watermark.py
• Lines: 1-111

Signature

class WMModel(ABC):
    @abstractmethod
    def get_watermark(self, x: torch.Tensor, message: tp.Optional[torch.Tensor] = None) -> torch.Tensor: ...
    @abstractmethod
    def detect_watermark(self, x: torch.Tensor) -> torch.Tensor: ...
    @property
    @abstractmethod
    def nbits(self) -> int: ...

class AudioSeal(WMModel):
    def __init__(self, generator, detector, nbits: int = 0): ...

Import

from audiocraft.models.watermark import AudioSeal, WMModel

I/O Contract

Inputs

Name	Type	Required	Description
x	torch.Tensor	Yes	Audio signal [B, C, T]
message	torch.Tensor	No	Message bits to embed [B, nbits]

Outputs

Name	Type	Description
watermark	torch.Tensor	Additive watermark signal [B, C, T] (from get_watermark)
detection	torch.Tensor	Detector output [B, 2+nbits, T] (from detect_watermark)

Usage Examples

from audiocraft.solvers.watermark import WatermarkSolver

# Load from checkpoint
model = WatermarkSolver.model_from_checkpoint('path/to/checkpoint', device='cuda')
watermark = model.get_watermark(audio, message=message)
watermarked = audio + watermark
detection = model.detect_watermark(watermarked)

Related Pages

• Principle:Facebookresearch_Audiocraft_Audio_Watermark_Training