DeepMark

DeepMark: Live Watermarking and Authentication for AI Voice

DeepMark embeds an imperceptible, machine-readable identity directly into AI-generated audio, image and video, so the content can later be detected as AI-generated and traced to its source. For voice agents, the mark is added to live speech and survives VoIP, codecs and the phone network, letting a bank, a contact center or the person on the line verify in real time that the caller is an authorized agent.

Key Features

• Voice agent authentication: Embeds a verifiable, authorized-agent identity into live AI speech that resolves to a trust record, so the receiving side can check who is calling.
• Phone-network robustness: The watermark is designed to survive VoIP, codecs, compression, transcoding and noisy phone channels.
• Imperceptible by design: The voice sounds the same and images and video look the same after marking.
• 128-bit payload: Carries up to 128 bits, enough for an agent or session identity, a model or provider ID, or a certificate pointer.
• Low compute cost: Under 2 ms of compute to watermark one second of audio, around 30,000 seconds of audio per second of GPU time on an L40S.
• Detection from short samples: Returns a confidence score and a binary watermarked or not-watermarked decision from brief audio clips.
• Multi-media marking: Marks AI-generated audio, image and video, not only voice.
• Vigil SDK: A C/C++ library for real-time streaming embedding and offline or real-time detection, with a C API for Python, Rust, Unity and other languages.
• Flexible deployment: Available as SDK, REST API, container, customer cloud, or on-prem and VPC, so content never has to leave your infrastructure.

Use Cases

• Banking and financial services: Let customers and staff confirm that an outbound AI call really comes from the bank, cutting the room for voice fraud.
• Contact centers: Mark every AI agent call so partners and callers can verify it in real time.
• Healthcare and insurance: Add a verifiable identity to automated patient and policyholder calls.
• AI content compliance: Mark synthetic media in a machine-readable format to meet EU AI Act Article 50 transparency obligations.

Getting Started

• Website: https://www.deepmark.me/
• Documentation: https://www.deepmark.me/docs/vigil/overview

DeepMark gives AI voice a built-in proof of origin: a watermark inside the audio itself rather than in metadata, so voice agents can be verified on live calls and synthetic content stays traceable after it is shared.