There is a subtle architecture shift happening in voice AI. The voice stack is becoming part of the agent's execution loop. Cartesia is combining the listening and speaking paths around that loop. Sonic-3.6 turns text into speech (90ms latency) and Ink-2 turns speech into text (100ms transcript latency), faster... than anything else streaming. now holds the #1 spot for both speaking and listening models. Becasue, voice agents are one of those systems where 100ms in the wrong place is very noticeable. Cartesia is attacking both sides of that loop at once. super interesting work by Karan Goel and team.show more

@cartesia tbh i haven't tried cartesia yet. is the 90ms consistent or does it spike under load

@cartesia Sub 100ms turnaround finally makes interruptions feel natural

@cartesia Voice makes latency a benchmark you can feel, and sub-100ms streaming just became the price of admission.

@cartesia 90ms latency is insane

@cartesia Sub-100ms latency is brutal, but the hidden trade-off is that streaming speed often comes at the cost of prosody and emotional range. the real edge in voice agents isn't just speed - it's making the machine sound less like a machine.

@cartesia Sp we get free upgrades to business and flight changes if we use their platform? Sold!

@cartesia In voice agents, latency stops being a benchmark number once it starts shaping the conversation.

@cartesia Latency is the whole game on live calls. People forgive a slightly synthetic voice much faster than they forgive a 700 ms pause. The pause is what makes callers say 'hello?' and talk over the agent. Shaving 100 ms off each side makes the interruption problem much smaller.

@cartesia the 100ms threshold is such a useful framing. once listening and speaking share the execution loop, turn-taking becomes a systems problem, not just a better tts demo.

@cartesia Shared listening-speaking loops could reduce latency beyond isolated voice models.

@cartesia 音箱刚更新俩新技术想想都激动。

@cartesia Voice AI is becoming part of the agent loop, not just an output layer. Huge shift for real-time agents.

@cartesia Sonic-3.6 and Ink-2 really are #1 on both arenas right now. The latency numbers are vendor-stated model latency though, not full round-trip, still impressive, just a nuance worth knowing. love it man

@cartesia This is pretty cool! It's exciting to see how voice AI is evolving and becoming more integrated.

@cartesia It is wild how much the latency gap is closing. Integrating the voice stack directly into the agent loop really feels like the missing piece for natural interaction.

@cartesia #1 on their own leaderboard. I don't think so.

@cartesia 100ms is the difference between a voice agent feeling turn-based and feeling present; putting both speech directions inside the execution loop matters more than another benchmark point.

90ms is fast enough that the bottleneck stops being the model and becomes everything around it — sip routing, barge-in detection, tool latency. saw a great postmortem this week: only 15% of build time on a phone agent went to conversation quality, the rest was telephony and failure handling. speed just moves where the real cost sits

@cartesia 90ms on the tts end isnt where a conversation feels slow. most of the gap is endpointing, waiting long enough to be sure the person actually stopped, and thats a few hundred ms you cant just cut without talking over people