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.

​
AudioSeal

Efficient localized audio watermarking for AI-generated speech with state-of-the-art robustness and ultra-fast detection

Get StartedView on GitHub

​
What is AudioSeal?

AudioSeal is a novel audio watermarking system designed for AI-generated speech. It uses localized watermarking at the sample level and a perceptual loss function to embed imperceptible watermarks that can be detected even in edited or compressed audio. The system consists of two neural network components:
  • Generator: Embeds watermarks into audio signals with minimal quality impact
  • Detector: Identifies watermark fragments in long or edited audio files with exceptional speed

Sample-Level Precision

Watermarks at 1/16,000 second granularity

Ultra-Fast Detection

2 orders of magnitude faster than existing models

Robust & Reliable

Survives compression, re-encoding, and noise

​
Key Features

Localized Watermarking

Watermark individual samples at 1/16,000 second intervals for precise tracking and detection

Minimal Audio Impact

Imperceptible watermarks that preserve the original audio quality

Streaming Support

Real-time watermarking for live audio applications with low latency

Secret Messages

Optional 16-bit message embedding for model version tracking or metadata

​
Quick Example

from audioseal import AudioSeal

# Load generator and detector models
generator = AudioSeal.load_generator("audioseal_wm_16bits")
detector = AudioSeal.load_detector("audioseal_detector_16bits")

# Watermark your audio
watermark = generator.get_watermark(audio_tensor)
watermarked_audio = audio_tensor + watermark

# Detect watermarks
result, message = detector.detect_watermark(watermarked_audio)
print(f"Detection probability: {result}")

​
Get Started

Installation

Install AudioSeal via pip or from source

Quickstart Guide

Embed and detect your first watermark in minutes

Core Concepts

Learn how AudioSeal’s watermarking technology works

API Reference

Explore the complete API documentation

​
Research & Resources

AudioSeal was developed by Meta’s Fundamental AI Research (FAIR) team and published at ICML 2024.
AudioSeal is available under the MIT license and can be used in both research and commercial applications.