Ink-2
Cartesia · USA · 2026
Transcription that also knows when you stopped talking — the listening half of Cartesia's voice loop.
Ink-2 is Cartesia's streaming speech-to-text model, announced together with the Sonic 3.5 voice and built on the same state space architecture the company uses instead of transformers. It transcribes continuously with a stated 100-millisecond transcript latency, but the feature that matters for voice agents is native turn detection: the model decides for itself when the speaker has finished a turn rather than leaving that judgement to a separate silence timer bolted on afterwards. That distinction is the difference between an assistant that waits awkwardly through a pause and one that interrupts before the sentence is over — the single most common reason people find AI phone systems unbearable. Paired with Sonic 3.5, Ink-2 gives Cartesia both directions of a real-time conversation from one vendor and one architecture, which is the company's core commercial argument against assembling a voice stack from separate transcription and synthesis suppliers.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!