Cartesia is an AI lab that builds real-time voice models and voice agents. Its models run on state space models (SSMs), a neural network architecture for long sequences such as audio that keeps a compact running state instead of comparing every part of the input with every other part, which keeps latency low. Its team has worked on SSM architectures such as Mamba.
Sonic is the text-to-speech model; it streams the first audio in about 90 milliseconds. Ink is the streaming speech-to-text model, with built-in turn detection that tells when a speaker starts and stops. Voice cloning copies a voice from a 10-second clip into more than 40 languages and accents. Managed Agents combines Ink, a language model chosen by the customer and Sonic into voice agents for phone calls, and handles turn-taking, tools and telephony. The models run in the cloud through regional API endpoints, on-premise or on-device.
Cartesia is a hosted API with a free plan and monthly subscriptions that include credits for the speech models; voice agent calls are billed per minute.