Cartesia

Cartesia builds speech models for conversation. Its Sonic text-to-speech line is engineered around latency — first audio in tens of milliseconds — because in a live phone call or voice assistant, delay is the difference between a conversation and an awkward pause.

What they provide

• Sonic TTS with very low time-to-first-audio, streaming output, and natural prosody
• Ink STT for fast, accurate transcription on the input side of the same loop
• Voice cloning from short samples, plus a library of designed voices
• Voice changing and style control for adjusting emotion, pace, and delivery
• On-device options for latency-critical or privacy-sensitive deployments

Why it matters technically

The models are built on state-space architectures rather than conventional transformers, which is what makes the latency and efficiency profile possible. For teams building voice agents, that shows up as agents that can interrupt and be interrupted naturally.

Access is through an API with a free tier for prototyping and usage-based pricing after that. Typical users are building phone agents, real-time translation, in-car and in-device assistants, and any product where a person is waiting to hear a reply.