Cartesia

Sweet spot: a developer team building an actual real-time voice agent — phone, app, kiosk — where the user is in a back-and-forth conversation with an AI. Latency is a perceptual cliff in voice UX, and Sonic is one of the few TTS models genuinely engineered for the streaming case rather than batch generation. Failure modes. If you are generating audio offline (audiobooks, video narration, podcast intros), Sonic's streaming-first architecture is wasted optimisation; ElevenLabs or Play.ht give more expressive options for that workload at similar cost. Voice cloning legal exposure is non-trivial — if you clone a real person without ironclad consent, the company you're building is what gets sued, not Cartesia. What to pilot. Wire Sonic into your real prototype with a real STT (Deepgram, AssemblyAI) and a real LLM, and measure end-to-end perceived latency from user-finishes-speaking to AI-starts-speaking. Below 800ms feels conversational; above 1.5s feels broken. If Sonic gets you under 800ms in your stack, commit; if not, the bottleneck is somewhere else (network, LLM TTFT) and changing TTS won't fix it.

Cartesia is a voice-AI company building Sonic, a state-space-model-based text-to-speech engine optimised for real-time, sub-100ms latency streaming. The pitch to developers is simple: if you are building a voice agent, IVR replacement, or any product where the user is talking to an AI in real time, traditional TTS (ElevenLabs, Google, Azure) introduces latency that breaks conversational feel. Sonic is architected from the ground up for streaming. The product surface is a clean REST + WebSocket API with SDKs in Python, TypeScript, Go, and Rust, plus a browser playground. It supports 40+ languages, instant 10-second voice cloning, professional voice cloning for commercial use, a curated library of pre-built voices, and inline emotion / laughter expression tags. SOC 2 Type II, HIPAA, and PCI Level 1 compliance are in place — uncommon for the open-research-flavoured TTS space. Cartesia has become the default voice layer for several real-time AI agent platforms (LiveKit, Pipecat, Retell, Vapi) and is available natively on Together AI's GPU clusters. For developer teams shipping voice agents in production, it is currently the strongest combination of latency, quality, and reliability.