DeepMark

AI-generated voice and media are increasingly used to impersonate people, including in scam calls to banks, contact centers, and insurers, making it hard for recipients to know if a call or piece of content is legitimate or AI-generated. Regulations like the EU AI Act's Article 50 now require synthetic audio, image, and video content to be marked in a machine-readable format and detectable as AI-generated, creating a compliance burden for providers. Existing approaches to watermarking face tradeoffs between latency, audio quality, robustness, capacity, and detection accuracy, and signals often fail to survive real-world conditions like VoIP, codecs, compression, and phone networks. Without a reliable way to embed and verify identity in generated media, businesses and customers have no practical way to confirm a call or piece of content is authentic.

DeepMark embeds an imperceptible, machine-readable identity signal directly into AI-generated audio, image, and video at the point of generation, which can later be detected and verified with clear pass/fail results from as little as a few seconds of audio. The flow has four steps: register an identity or credential, embed the signal during generation without degrading quality, detect and verify the signal later in a call, API, or network, and resolve the payload into a provable result backed by signed evidence. The system supports up to 128 bits of payload capacity, operates in under 2 ms of compute per second of audio, and is built to survive real-world degradation including VoIP, transcoding, codecs, compression, cellular networks, and re-recording. It deploys via SDK, REST API, container, customer-cloud, or on-prem/VPC, so sensitive content can be marked and verified inside a customer's own environment without leaving their network.

The EU AI Act's Article 50 requires synthetic audio, image, and video to be marked in a machine-readable format and detectable as AI-generated, creating a near-term compliance obligation DeepMark's marking and detection layer is built to support.