DeepMark, a five-person startup in Y Combinator's Fall 2026 batch, has launched a live watermark that embeds machine-readable IDs in AI-generated speech and, the company said, holds up over phone calls. The debut arrives as OpenAI, Google and European regulators push watermarking toward the mainstream, and as contact-center data suggests AI-assisted fraud has stopped being a curiosity.
The YC listing describes the product as "a live watermark inside AI voice that survives phone calls." DeepMark said the system marks audio in real time for AI voice agents. Detection, it said, can run on a phone or in a call center, or through an API.
Who Is Behind It
Slavko Kovačević, the chief executive, and Kosta Pavlović, the chief science officer, founded the company in 2025. It is headquartered in San Francisco, registered in Delaware and has a presence in Podgorica, Montenegro. According to the company's page, both founders are researchers in learning-based audio watermarking, and their papers have been cited by Microsoft, Meta, Sony and Adobe. That claim comes from DeepMark itself, and I could not tie it to a specific citation record.
A Crowded, Converging Field
The largest platforms are settling around Google's SynthID. Google DeepMind opened a public SynthID Detector portal on May 20, 2025. In a July 31, 2026 update, OpenAI said that "supported audio generated with OpenAI tools, including through ChatGPT and the OpenAI API, now includes SynthID watermarking." ElevenLabs has said it is working with Google DeepMind to put SynthID into its generated audio.
Some players got there sooner. Meta FAIR published AudioSeal, a localized audio watermark, in June 2024 and released the code as open source. Resemble AI introduced its PerTh neural speech watermark in 2023, and the company made watermarking the default in 2026.
What Is Pushing Demand
Fraud data is the most immediate pressure. Pindrop's June 12, 2025 report said deepfake fraud attempts rose by more than 1,300% in 2024, and that contact centers could face $44.5 billion in fraud exposure in 2025. More recently, Pindrop's Q1 2026 figures, drawn from its own customer dataset, put AI-backed attacks at 8.7% of all contact-center fraud, up from under 1% a year earlier. Because the sample is Pindrop's customers rather than the industry at large, the number is best read as a signal, not a census.
Regulators are adding their own push. Under the EU AI Act, Article 50 transparency obligations applied from Aug. 2, 2026. Providers whose systems were already on the market before that date have until Dec. 2, 2026 to meet the marking and detection requirements in Article 50(2).
In the United States, the FCC ruled in February 2024 that the Telephone Consumer Protection Act's "artificial or prerecorded voice" language covers AI-generated voices. "We're putting the fraudsters behind these robocalls on notice," Chairwoman Jessica Rosenworcel said at the time.
Then there is the business case. Gartner predicted in March 2025 that agentic AI will autonomously resolve 80% of common customer service issues by 2029. Yet a Gartner survey published Aug. 4, 2026 found that 87% of customers say companies using generative AI for service must also offer access to a human agent. Companies, in other words, are being asked to automate and to be transparent about it, which is where provenance tools find an audience.
Why Phone Lines Are Hard
DeepMark's launch post makes the technical case plainly: "A watermark that works beautifully in the lab falls apart the moment audio meets a phone line."
Live calls are typically transmitted as narrowband, 8 kHz–sampled speech and run through telephony codecs, which can degrade or remove the subtle signal a watermark relies on. An academic project called CallShield, posted to arXiv on Jan. 14, 2026, targets real-time watermarking of live 8 kHz speech and embeds bits in 40 ms frames.
Existing caller ID defenses do not fill the gap. A Congressional Research Service report said STIR/SHAKEN mitigates caller ID spoofing but does not authenticate the voice or the content of a call.
The Evidence, and Its Limits
DeepMark's research record is its main public proof. Its benchmark paper, "Redefining Audio Watermarking Robustness," appeared in IEEE Access 14 in 2026, and its AWARE paper was accepted at EUSIPCO 2026. A zero-bit backward-compatibility paper followed on arXiv on Oct. 19, 2026.
Commercial traction is harder to verify. The YC page says the company's first platform license is "now in production." DeepMark has not named the customer, and none had been publicly identified as of Nov. 2, 2026.
Outside work complicates the picture. A Sept. 29, 2026 arXiv paper found that speech enhancement chains can remove or impair neural audio watermarks, and it tested AudioSeal, WavMark, SilentCipher, Timbre, Perth and AlignMark. A separate Aug. 15, 2026 paper proposed training-free removal methods. It found that latent-domain marks such as AWARE were more robust than carrier-domain ones under the conditions tested. AWARE is DeepMark's own work, which is encouraging for the company, though hardly independent validation of its phone-call claim.
Detection vendors sit on adjacent ground. Pindrop, Reality Defender, Hive and Aurigin.ai sell audio deepfake detection and do not embed marks. Research cited by Microsoft Research in January 2026 treats provenance, watermarking and fingerprinting as separate pillars rather than interchangeable fixes.
Where This Leaves DeepMark
Estimates for the broader voice security market vary widely. Fortune Business Insights values voice biometrics at $2.87 billion in 2025, rising to $22.76 billion by 2034. Mordor Intelligence puts it at $2.63 billion in 2025 and $6.54 billion by 2031. Neither figure measures watermarking directly, so they are only a rough guide to the neighborhood.
The SynthID ecosystem has a limit of its own. Its detectors often work only within partner ecosystems, so a mark from an outside voice agent may not be readable by them. DeepMark is pitching itself at that gap.
The company said on its page that it is expanding into image and video provenance, and into voice-agent authentication alongside telecom fraud prevention. Kovačević struck a sober note when he spoke to the Montenegro Innovation Fund in October 2026. "The work will not be done by itself," he said, "and we now expect more challenges and effort than ever."
