Cartesia lands $64M Series A and rolls out Sonic 2.0

Nothing ruins the illusion of talking to an AI quite like that dead air after you finish speaking. That awkward one-second pause is usually the difference between a real conversation and waiting for a machine to buffer.

Cartesia wants to eliminate that gap entirely. The real-time voice startup just raised a $64 million Series A led by Kleiner Perkins, and to celebrate the check, it rolled out its newest flagship model: Sonic 2.0.

Cutting latency down to 40 milliseconds

When it comes to spoken conversation, human turn-taking typically happens in about 200 milliseconds. If an AI takes half a second to start generating words, natural banter falls apart.

Cartesia claims Sonic 2.0 drops latency down to 90 milliseconds for its full models, and an astonishing 40 milliseconds in optimized real-time setups. For teams pushing the limits even harder, they also launched an ultra-fast tier called Sonic Turbo.

The secret sauce here is the architecture. Instead of sticking strictly to standard transformers, Sonic 2.0 leverages state space models (SSMs). Cartesia doubled the model size compared to the previous generation, yet managed to increase both throughput and efficiency.

It is not just about raw speed either. The new model steps up voice cloning precision, capturing nuanced accents, stylistic inflections, and tonal shifts. That makes it instantly relevant for tricky applications like automated support call centers, accent-accurate localization, and accessibility tools.

Built for enterprise scale and edge devices

Cartesia is clearly courting enterprise buyers right out of the gate. Alongside the model drop, the company highlighted a 99.9% uptime track record and confirmed compliance with both SOC-2 and HIPAA standards.

That enterprise readiness allows developers to plug the Sonic API directly into medical voice transcription or legacy corporate phone banks without running into legal roadblocks.

Crucially, Cartesia is offering on-device deployment options alongside its cloud API. CEO Kar stated that this is officially the year of voice AI, noting that synthetic voices are moving rapidly into everyday enterprise infrastructure. With the new capital, the team is working on streaming optimizations, on-device inference, and upcoming features like voice morphing and audio infill editing.

Why it matters

We think Cartesia's strategy focuses on the exact pain point holding conversational agents back: latency. While big tech companies battle over reasoning benchmarks, voice remains the most direct human interface, and whoever owns sub-100ms conversational audio holds massive leverage.

For developers building call center agents, voice translators, or physical companion devices, Sonic 2.0 offers an immediate performance upgrade worth testing today. Getting genuine 40ms audio response times allows real-time interruptions and back-and-forth cadence that felt impossible a year ago.

Offering on-device inference alongside cloud APIs is particularly clever. It shields developers from spiraling server bills and solves strict privacy headaches, putting direct pressure on market favorites like ElevenLabs to speed up their turnarounds.

That said, banking on state space architectures still carries real-world execution risks. While SSMs deliver incredible speed benefits on paper, proving long-term stability and expressive vocal range across millions of unstructured enterprise edge cases is Cartesia's next big test.