by CartesiaUS
Make AI phone and voice agents sound and react like a person, answering in under a tenth of a second, so talking to software stops feeling like waiting on a machine.
Sonic-3.6 out Aug 2026; #1 on Artificial Analysis voice boards; 61 locales.
Updated 27 Aug 2026Checked 11 Oct0 updates this week
Audio is generated piece by piece and streamed, so the first syllable plays in under 90 ms while the rest is still being made.
State space layers carry a compact summary forward instead of re-reading all prior audio, keeping each step cheap.
Given a brief recording, the model reproduces that voice and accent; emotion and pace can be steered per line.
Ink transcribes the caller and spots the end of a turn from meaning, not silence; Sonic then speaks the reply.
| Spec | Sonic | Sonic-3.5 | Change |
|---|---|---|---|
| Time to first audio | Under 90 msR (reported) | — | —no comparable change |
| Generation speed | About 132 characters/sR (reported) | 68 (v3 Conversational) | ▲ 64 |
| Locales | 61 (11 Indic)R (reported) | — | —no comparable change |
| Preset voices | 500+R (reported) | — | —no comparable change |
| Artificial Analysis Elo (controlled) | 1120, ranked #1R (reported) | 1095 (Sonic-3.5) | ▲ 25 |
| Blind test vs Eleven v3 (US English) | 92% preferred Sonic-3.6R (reported) | — | —no comparable change |
R reported by the company
Cartesia
Real-time voice models (Sonic TTS, Ink STT), state space models
Cartesia builds very fast speech models for live voice agents: Sonic turns text into speech and Ink turns speech into text. Founded by Stanford researchers behind state space models, an alternative to transformers.
