Cartesia’s new voice models hit 90ms latency and detect when you’re done talking

AI voice startup Cartesia just dropped two new models: Sonic-3.5 for text-to-speech and Ink-2 for speech-to-text, bundled into a real-time voice agent stack.

Sonic-3.5 is built for low-latency voice generation, with a time to first audio of just 90 milliseconds. It ships with support for 42 languages out of the box and handles English heteronyms and alphanumeric pronunciation without preprocessing.

Ink-2, on the other hand, brings the word error rate down to 3.6% and introduces native turn-detection — a noise-aware mechanism that figures out when you’ve stopped talking based on semantic understanding rather than just silence duration. For now, Ink-2 only supports English; multilingual versions are coming later.

Developers can call both models through a single API. Sonic-3.5 and Ink-2 are designed for bidirectional streaming to cut the latency and overhead you’d get from stitching together different vendors.