Cartesia Review 2026

Real-time generative voice and on-device speech models built on state-space architectures — Sonic TTS at ~40ms first-token latency, Ink-Whisper STT, voice cloning, and an Edge SDK for offline voice on devices.

Cartesia is a San Francisco AI lab founded by the authors of the Mamba and S4 state-space model papers. Its bet is that state-space architectures are a better foundation than transformers for streaming audio, and the products bear that out. Sonic is Cartesia's flagship text-to-speech model with first-token latency around 40 milliseconds, making it one of the fastest production TTS systems and well-suited to live voice agents where any delay degrades conversational feel. Sonic supports multilingual output, instant voice cloning from short audio clips, professional voice cloning for studio-quality custom voices, and emotion / style controls. Ink-Whisper is Cartesia's STT model optimized for the same low-latency budget. The Edge SDK runs compressed versions of Cartesia models directly on phones and embedded devices for offline voice. Pricing: Free Hobby plan ($0 with monthly credits), Pro at ~$1/mo for testing, Scale at $299/month for production volumes, and Enterprise (custom) — with usage-based per-minute charges layered on top (e.g., ~$0.06/minute for Sonic, rates varying per model). Cartesia is the default low-latency TTS choice in many voice-agent stacks built with LiveKit, Pipecat, Vapi, and bespoke WebRTC pipelines.