RESEMBLE WATERMARKER

Multimodal, imperceptible watermarks for audio, image, video and text

Embed proof of ownership, authenticity and verifiable provenance into every audio, image, video and text file. Built to support EU AI Act Article 50 compliance.

Contact us
Try it Free
Trusted by

Watermarking models for audio, image, video and text

The original PerTh model is open source, PerTh Multimodal is available to enterprise customers.

PerTh Multimodal
VERIFY
Enterprise and compliance-oriented multimodal model, extending PerTh's imperceptible, tamper-resistant watermarking across audio, video, image, and text in a single API. Each modality uses a purpose-built algorithm, with robustness tuned against real-world transformations. Supports explicit marks (custom encodes) for audio, image, and video.
Audio, image, video and text
Real-time encode and decode
API available
EU AI Act Article 50 oriented
GET STARTEDLEARN MOREDOCSBENCHMARKS
PerTh
VERIFY
A deep neural network that embeds imperceptible, psychoacoustically-masked watermarks into audio at the point of creation. Survives MP3 compression, audio editing, noise, and codec transforms. Embedded in every Chatterbox output by default. MIT-licensed and available as open source.
99.9% decode accuracy
Survives compression
Audio-only watermarking
Open source and free
LEARN MOREGITHUBDocs
WHAT A MARKED ASSET PROVES

The four questions a watermarked asset answers

Encode a mark at creation, decode it later, and a single asset can answer the questions that ownership and compliance actually turn on:
Who issued this content?
Was it AI-generated, AI-altered, or authentic?
Has it been tampered with since creation?
Are other provenance marks present, including SynthID or C2PA credentials?
Resemble Watermarker reads SynthID and C2PA markers, and writes and signs C2PA which creates a provenance manifest which supports EU AI Act Article 50 compliance.
Try Resemble Watermarker free
How WATERMARKING WORKS

Encoded at creation, decoded on-demand

Encode once, and the proof travels with the file. Watermarks remain part of the file itself and ready for decode any time proof of compliance is required. Three steps:
STEP 1 - ENCODE
Signal encoding in the file
Any asset (audio, video, image, or text) is watermarked at creation - or at the time of your choosing - via a single API call. Encode a Resemble Watermark and optionally sign with C2PA. Encode an explicit watermark (custom message) for complaince or tracking purposes.
STEP 2 - TRAVEL
Distribute marked content
The mark survives compression, re-encoding, editing, and format conversion across all four content types. Distribute content with the confidence it carries proof of ownership and a declaration of the presence of AI.
STEP 3 - DECODE
Decode to read results
Submit any downstream file to recover the mark, confirm presence, and read its attribution - including SynthID, C2PA and Resemble Watermarks. Verification always requires this decode step, marks are machine-readable and imperceptible to humans.
WHAT AI WATERMARKING PROVES

A watermark proves who made a piece of content and whether it's been changed

Metadata is easily removed and often lost
Every platform strips it on upload. By the time you need to prove ownership, it's already gone. An in-signal watermark stays with the content through compression, re-encoding, and re-uploads.
A birth certificate and fingerprint for content
 Deepfake detection provides probabilistic scoring on whether AI was used to create or alter content. A watermark reads back who issued it, whether it's AI-generated or altered, and whether it's changed since creation - declared provenance.
Regulation is starting to require watermarking
The EU AI Act's Article 50 requires AI-generated audio, video, image, and text to carry machine-readable marking, effective August 2, 2026. One of the key requirements is a mark that declares not just ownership, but whether content is AI-generated or altered with AI.
built to survive

Imperceptible encodes that survive re-encoding, compression, and format conversion

Audio watermarking
WHAT IT SURVIVES

Noise addition, pitch shifting, reverb, time-stretching, format conversion

Audio watermark is trained on real-world manipulations, maintaining ~100% detection accuracy even after significant processing.

HOW IT WORKS

Constrained to speech-relevant frequencies — inaudible to listeners

Embedded in the signal itself, not in file metadata making removal without audible corruption nearly impossible.

OUTPUT

Decode — watermark and C2PA presence

Read the presence of Resemble Watermarks, including 16-bit explicit marks. Decode for the presence of C2PA and SynthID.

Image + video watermarking
ATTACK TYPE

Screenshots, re-encoding, compression, platform reprocessing, format conversion

The watermark is embedded in the visual data itself, not in EXIF, XMP, so it survives every transformation that strips metadata.

DETECTION METHOD

Auto-detected media type — one endpoint for image and video

Submit any file. Resemble detects and processes images or videos and applies the appropriate watermark. No configuration per file type required.

OUTPUT

Binary decode after any point in the content lifecycle

Decode post-screenshot, post-repost, post-platform-reprocessing. Explicit marks are 256-bit for video, custom messages can hold all relevant information.

C2PA + provenance
WHAT IT PROVIDES

Second provenance layer on top of the embedded watermark

C2PA (Coalition for Content Provenance and Authenticity) enrollment creates a cryptographically signed metadata manifest linked to the watermark signal.

HOW IT WORKS

Automatic C2PA enrollment on every watermark — no extra API call

Every file is enrolled in C2PA by default. If the manifest is stripped, the embedded watermark acts as a fallback provenance signal.

OUTPUT

Machine-readable AI origin — satisfies EU AI Act Article 50

C2PA manifest declares AI origin in a standardized, machine-readable format, satisfying compliance beginning August 1, 2026.

AI CONTENT RISK BY INDUSTRY

Built for regulated industries

Where unverified AI content carries legal, clinical, or financial consequences.

Banking & Finance
Industry risk:
Voice-clone fraud bypassing biometric systems.
Resemble AI solution:
Prove a disputed transaction used an authentic voice, not a synthetic clone.
Law Enforcement
Industry risk:
Evidence excluded without documented provenance.
Resemble AI solution:
Establish chain of custody for audio, image, and video evidence.
Telecommunications
Industry risk:
FCC non-compliance on AI-generated voice calls.
Resemble AI solution:
Prove which calls originated from your systems at scale.
Media & Entertainment
Industry risk:
Unauthorised use of synthetic talent voices.
Resemble AI solution:
Make every generated file self-evidencing for talent rights and IP.
Healthcare & HealthTech
Industry risk:
Unverified synthetic instructions in clinical pathways.
Resemble AI solution:
Verify clinical voice instructions are authentic before they reach care teams.
Generative AI platforms
Industry risk:
EU AI Act compliance, mandatory from August 2026.
Resemble AI solution:
Watermark content by default at point of generation. EU AI Act ready.
SOC2 Type II
Enterprise plans
EU AI Act ready
Mandatory Aug 2026
GDPR Compatible
On-prem available
HIPAA Compatible
Air-gapped deployment
SSO / SAML
Enterprise identity
C2PA Standard
Content provenance
AI SECURITY

Trusted where it matters most

When the stakes are real, rely on Resemble AI
Case studiesContact us
U.S. Senate Judiciary
Resemble AI's CEO testifies on election deepfakes, presenting detection infrastructure as a national security requirement.
Carahsoft Public Sector
Available to US Federal, State and Local Government across hundreds of contract vehicles.
Canada AI Code of Conduct
Signed alongside IBM and Scale AI, as a pillar of responsible AI deployment.
AI narration for The Andy Warhol Diaries
Generated from three minutes of source audio, PerTh watermarked every file for talent IP protection and authenticity disclosure.
INTEGRATIONS AND DEPLOYMENTS

Go live in hours, not sprints.

Rest API
Python SDK
Node.js SDK
JavaScript SDK
MCP for Cursor & Claude Code
On-prem available
View API Docs
View GITHUB
Frequently asked questions
Does watermarking affect audio or image quality?
No. PerTh embeds into inaudible frequencies for audio, imperceptible pixel variations for images and video, and subtle linguistic changes for text. Quality stays at production grade across all four media types, and end users notice nothing.
What is C2PA and how does it work with PerTh watermarking?
C2PA is an open standard for embedding a content provenance manifest into a file's metadata. It tells platforms and regulators who created the content and how it was produced. PerTh works alongside C2PA: C2PA handles the metadata layer, PerTh embeds directly into the file itself. If the metadata is stripped, the watermark remains. Both layers together provide the most complete provenance record available.
Can we detect if our data was used to train another model?
Yes. PerTh Multimodal's watermark persists through AI model training pipelines. If your watermarked content was used without authorization in a third-party training run, the signal survives and can be verified. This provides a direct line of defense against unauthorized IP use.
Does my data leave our infrastructure during API processing?
On-premise and air-gapped deployment options mean media files never leave your own infrastructure. For teams using the cloud API, contact us for specifics on processing location and compliance coverage for GDPR and HIPAA.
Is Resemble Watermarker compliant with the EU AI Act?
Yes — it's built to be compliance-oriented for Article 50, which requires machine-readable marking of AI-generated content and takes effect August 2, 2026. PerTh Multimodal produces machine-readable marks across audio, video, image, and text, and reads and writes C2PA, mapping to the Commission's "effective, interoperable, robust, reliable" bar. Non-compliance carries fines up to €15 million (or 3% of worldwide annual turnover).
How quickly can we integrate?
Most teams are live within a few hours via REST API. Python, Node.js, and JavaScript SDKs are available. MCP server support for Cursor and Claude Code is included. Synchronous mode returns results in a single request without polling.
Get complete generative AI security
Join thousands of developers and enterprises securing with Resemble AI
Contact Us