Cartesia

Cartesia was founded by the researchers behind state-space models — the Mamba architecture — and it shows in the one number that matters most for live voice: latency. Its Sonic 3.5 text-to-speech regularly tops real-time voice leaderboards, and the companion Ink-2 transcription model rounds out a stack purpose-built for AI phone agents and interactive assistants where any lag breaks the illusion of a conversation. For developers, the appeal is a single team and API handling both speech-out and speech-in, plus voice cloning and a generous free tier (about 20k credits a month at $0) that makes it easy to prototype before the paid Pro/Startup/Scale tiers kick in. The honest caveats: pricing is credit-based, so you'll do a little arithmetic to map credits onto characters and cost; it's a voice specialist rather than an all-in-one AI platform; and its language and voice catalog is still smaller than ElevenLabs. But when your product lives or dies on how instantly a voice responds, Cartesia is one of the strongest options on the market — and worth putting head-to-head with ElevenLabs before you commit.