Audio AI Watermarking: OpenAI, Meta, and Google SynthID

OpenAI extended SynthID watermarking to audio on July 31, 2026. It covers supported audio from ChatGPT Voice and GPT-Live, plus the API.

Alongside it came a Content Provenance API, so third parties can check for OpenAI's provenance signals programmatically instead of relying on a manual verification page.

That date sits two days before EU AI Act Article 50 became applicable on August 2, 2026. OpenAI's own GPT-Live system card, dated July 8, 2026, does not mention SynthID, C2PA or watermarking anywhere in it.

Read together, that reads less like provenance designed into GPT-Live from the start and more like a compliance layer bolted onto a model whose safety documentation had already shipped.

Meta watermarked audio before it watermarked images under current branding. That ordering surprises almost everyone.

The Llama Defenders Program's Llama Audio Watermark Detector, also called the Llama Generated Audio Detector, shipped on April 29, 2025. Content Seal, the image system that got all the coverage, did not launch until July 7, 2026.

The audio detector was built for a narrow job: voice-fraud detection, not general AI-content labeling.

Its early adopters say the rest. Zendesk, Bell Canada and AT&T are companies that need to catch AI-cloned voices in call centers and telecom fraud, in real time.

That is a security problem, not a transparency one. Meta treated audio provenance as the former more than a year before it treated image provenance as the latter.

Google DeepMind holds US patent 12,094,474 B1, filed November 15, 2023 and granted September 17, 2024. A continuation was published May 8, 2025 as US 2025/0149048 A1.

It describes pairing a watermarking neural network, an encoder and decoder, with a separate embedding and fingerprint network, both trained adversarially against deliberate distortions.

The output is two scores, a watermark-detection score and an embedding-similarity score. Combined, they decide whether an image or audio object came from a specific generative system.

The named inventors include Sven Adrian Gowal, Christopher Gamble, Florian Nils Stimberg, Sylvestre-Alvise Guglielmo Rebuffi, and Ian Goodfellow, the researcher whose 2014 paper introduced the generative adversarial network in the first place.

Because OpenAI's own audio watermarking is built on Google's SynthID, this patent family plausibly sits underneath OpenAI's audio provenance signal too, at least indirectly.

Meta's separately built Llama Audio Watermark Detector has no documented connection to it at all.

Google does not treat audio as a separate product. Its own documentation describes SynthID as one underlying technology applied across images, audio, text and video.

Gemini's help page for verifying AI-generated content reflects that unevenly. It covers audio clips up to one hour long, alongside images and video, but stops short of covering text at all.

Synthetic speech carries a risk the other modalities don't carry in the same shape. Voice cloning enables direct impersonation over the phone, in voicemail and inside voice-authentication systems.

That is why EU AI Act Article 50 covers synthetic audio in both its provider marking duty and its deployer deepfake-labeling duty, and why telecom and support companies, not media platforms, were the first customers for Meta's detector.

Outside the EU, the rules genuinely disagree with each other:

• California's AI Transparency Act requires covered providers to support latent, metadata-level disclosure (provider name, system version, a timestamp, a unique ID) on AI-generated audio, image and video alike.
• One law firm's summary of India's new platform rules goes further on specifics, stating that a visible label must cover at least 10 percent of an audio clip's starting duration.
• New York's Synthetic Performer law, built to require disclosure when an ad uses an AI-generated human likeness, explicitly exempts audio-only ads from that duty.

Same technology, three different answers about whether a synthetic voice needs a label at all. New York's carve-out is the most revealing of the three, because it treats the fabricated face, not the fabricated voice, as the thing that warrants a warning.