Resemble AI is a useful case study in how fast this industry reorganises itself. It built its name on voice cloning, generating a synthetic voice from a modest sample, and it has repositioned substantially toward the opposite problem, detecting synthetic audio, video and images and establishing provenance for content. Anyone arriving from a search for voice cloning tutorials should know that up front, because the company you find is not quite the company the search results describe. The pivot makes commercial sense and is worth understanding on its own terms.
Voice generation has become crowded and heavily commoditised, with strong open models and aggressive price competition. Detection and provenance, meanwhile, are becoming procurement requirements rather than nice-to-haves, driven by fraud losses at contact centres, by executive impersonation attacks that have moved from theoretical to routine, by identity verification processes that now have to assume the submitted video might be generated, and by regulation arriving in several jurisdictions. A company that already understands how synthetic audio is produced is genuinely well placed to detect it, and that is a more defensible position than competing on voice quality. The published research is the strongest thing here and the main reason to spend time on the site.
The deepfake threat reporting, including a midyear review for 2026, and the incident database tracking real-world attacks, are substantive documents rather than gated marketing assets with a form in front of them. If you work in trust and safety, fraud prevention or security awareness, this material is useful for understanding what attacks are actually being attempted rather than what is theoretically possible, and that distinction matters when you are deciding where to spend a defensive budget. The incident tracking in particular gives you concrete cases to reason from. The detection benchmarks are published with specific numbers, figures in the high nineties for audio and slightly lower for video and images.
I want to be careful here. Published accuracy figures from a detection vendor, measured on benchmarks that vendor selected, are a starting point for a conversation and not a basis for a procurement decision. Detection is adversarial, which means performance against a benchmark assembled at one moment tells you relatively little about performance against generation methods that appear afterwards. Anyone evaluating this seriously should test against their own adversarial examples and should expect worse numbers than the published ones.
To Resemble's credit, publishing specific figures at all is better than the vague gestures at accuracy that some competitors offer. The watermarking work addresses a real and underappreciated problem. Establishing that a piece of content came from where it claims is going to matter more as generation gets cheaper, and technical provenance mechanisms are part of the answer even though they are not the whole answer, since watermarks can be stripped and absence of a watermark proves nothing. The documentation on this is clear about what the mechanism does.
On the documentation itself, it is competent developer reference material. API endpoints, SDK libraries, authentication, integration patterns. It tells a developer how to call the service. It does not teach anything conceptual about how detection works, what its failure modes are or how to design a workflow that degrades sensibly when a detector is uncertain, and that last question is the one that matters most in deployment.
A detector that returns a probability is only useful if somebody has decided what happens at each threshold, and that decision is a policy question the documentation does not touch. The pricing is quote-based with a free entry point advertised and no published tiers, which is standard for enterprise security tooling and still worth noting as a friction point for anyone trying to evaluate quickly. Three point four. Genuinely valuable public research, clear developer documentation, no structured education, and a company whose current product is not the one its search traffic is looking for.