From generating voices to inspecting media
Resemble AI's work spans generative voice and detection of synthetic media. Its Chatterbox voice models remain available with an MIT-licensed model card, giving developers a voice-generation component they can run outside a closed application. That research history remains relevant even as the company's business changes. [3]
In August 2026, its CEO said the company was no longer taking new paid voice-generation customers and had put its commercial effort behind detection. Existing customers would receive support, while Chatterbox and other voice research would remain available. Describing it simply as a current commercial voice-cloning service would miss that stated shift. [1]
Detecting signals, with explicit limits
The company released DETECT-World on August 12, 2026. It described combining known generator artifacts with learned physical-consistency signals, producing manipulation probabilities, per-frame scores and spatial heatmaps through an API. [2]
The release explicitly calls its score a probability rather than proof. It also distinguishes externally evaluated audio results from internal image and video benchmarks awaiting external validation. Those limits are important to the company's ecosystem role: helping reviewers inspect generated media, rather than certifying every recording as authentic or manipulated. [2]
Sources
- Why a former voice AI company went all in on deepfake detection
Resemble AI · Aug 13, 2026
- Introducing DETECT-World: The First World Model for Deepfake Detection
Resemble AI · Aug 12, 2026
- ResembleAI/chatterbox
Resemble AI · retrieved Oct 23, 2026