Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/facebookresearch/audioseal/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

The AudioSeal class provides static methods to load pre-trained generator and detector models. It serves as the primary interface for model initialization in AudioSeal.

​
Methods

​
load_generator

Load a pre-trained AudioSeal watermark generator model. 
from audioseal import AudioSeal

generator = AudioSeal.load_generator(
    model_card_or_path="audioseal_wm_16bits",
    device="cuda"
)

​
Parameters

​
model_card_or_path
str
required
Model card name or path to checkpoint. Can be:
  • A local model card name (e.g., “audioseal_wm_16bits”)
  • A path to a local checkpoint file
  • A Hugging Face model path (e.g., “facebook/audioseal/generator_base.pth”)
  • An HTTPS URL to a checkpoint
​
nbits
int
default:"None"
Number of bits for the watermark message. If not specified, uses the value from the model config.
​
device
Device
default:"None"
Device to load the model on (e.g., “cpu”, “cuda”, “cuda:0”). If None, defaults to CPU.
​
dtype
DataType
default:"None"
Data type for model parameters (e.g., torch.float32, torch.float16). If None, uses default precision.

​
Returns

​
generator
AudioSealWM
The loaded generator model ready for watermark generation.

​
Example

import torch
from audioseal import AudioSeal

# Load from model card
generator = AudioSeal.load_generator("audioseal_wm_16bits")

# Load with custom device and dtype
generator = AudioSeal.load_generator(
    "audioseal_wm_16bits",
    device="cuda",
    dtype=torch.float16
)

# Load from Hugging Face
generator = AudioSeal.load_generator(
    "facebook/audioseal/generator_base.pth"
)

# Load from local checkpoint
generator = AudioSeal.load_generator(
    "/path/to/checkpoint.pth",
    device="cuda"
)

​
load_detector

Load a pre-trained AudioSeal watermark detector model. 
from audioseal import AudioSeal

detector = AudioSeal.load_detector(
    model_card_or_path="audioseal_detector_16bits",
    device="cuda"
)

​
Parameters

​
model_card_or_path
str
required
Model card name or path to checkpoint. Can be:
  • A local model card name (e.g., “audioseal_detector_16bits”)
  • A path to a local checkpoint file
  • A Hugging Face model path (e.g., “facebook/audioseal/detector_base.pth”)
  • An HTTPS URL to a checkpoint
​
nbits
int
default:"None"
Number of bits for the watermark message. If not specified, uses the value from the model config.
​
device
Device
default:"None"
Device to load the model on (e.g., “cpu”, “cuda”, “cuda:0”). If None, defaults to CPU.
​
dtype
DataType
default:"None"
Data type for model parameters (e.g., torch.float32, torch.float16). If None, uses default precision.

​
Returns

​
detector
AudioSealDetector
The loaded detector model ready for watermark detection.

​
Example

import torch
from audioseal import AudioSeal

# Load from model card
detector = AudioSeal.load_detector("audioseal_detector_16bits")

# Load with custom device and dtype
detector = AudioSeal.load_detector(
    "audioseal_detector_16bits",
    device="cuda",
    dtype=torch.float16
)

# Load from Hugging Face
detector = AudioSeal.load_detector(
    "facebook/audioseal/detector_base.pth"
)

# Load from local checkpoint
detector = AudioSeal.load_detector(
    "/path/to/checkpoint.pth",
    device="cuda"
)

​
Notes

  • Model checkpoints are cached in ~/.cache/audioseal by default
  • You can override the cache directory using the AUDIOSEAL_CACHE_DIR environment variable
  • For Hugging Face models, the library will automatically handle authentication if needed
  • Loading gated models requires the huggingface_hub package installed

​
See Also

Build docs developers (and LLMs) love

Get started for freeTalk to us