Cartesia takes top spots for real-time speech and listening

If you have ever tried talking to a voice agent on the phone, the most frustrating part is rarely what it says. It is the agonizing delay before it speaks, or the way it interrupts the moment you pause to breathe.

Cartesia rolled out Sonic-3.5 and Ink-2, a paired voice release designed to fix that exact handoff problem. Sonic-3.5 handles text-to-speech, while Ink-2 takes speech-to-text. According to Cartesia, both models placed number one on Artificial Analysis's streaming leaderboards for naturalness and transcription accuracy.

Founder Karan Goel called it out directly: Cartesia is currently the only vendor holding the top streaming spot for both listening and speaking at the same time.

Cutting the latency tax

Most voice engines rely on traditional transformer stacks, but Cartesia built both models on a state space model architecture. That architectural detour is the main reason they can strip out so much latency.

For Sonic-3.5, Cartesia reports a median time to first token below 300ms. That speed keeps a back-and-forth conversation moving fast enough that neither party feels trapped behind a buffering icon. Compared to Sonic-3, the new model offers better rhythm, wider emotional range, and clearer pronounciation for alphanumeric IDs and codes.

Sonic-3.5 lists over 40 supported languages, though Cartesia notes in its changelog that non-English languages are still expanding from an English-first base. Multilingual teams will want to run their own evals first.

Meanwhile, Ink-2 handles streaming transcription in noisy real-world environments. Its biggest trick is built-in turn detection and semantic endpointing. Instead of relying on a crude silence timer, Ink-2 understands whether a speaker actually finished an idea or simply paused mid-thought.

Both models are now generally available via WebSocket and REST APIs, alongside a public playground, Python and TypeScript SDKs, and prebuilt integrations for Pipecat and LiveKit.

Why it matters

We think Cartesia's strategy here is sharp and targeted.

A conversational voice agent is a tight loop: transcribe, reason, synthesize, repeat. Historically, engineers had to patch together transcription from one vendor, reasoning from another, and voice output from a third. Every hop added network latency and increased the odds that the agent would talk over someone or sit in dead silence.

By owning both the ear and the mouth of the stack, Cartesia can optimize the exact moment the turn-taking handoff occurs. For anyone building AI phone representatives, live language tutors, or interactive assistants, this dual release removes a ton of painful orchestration work.

The voice landscape is clearly splintering into niches. If you need rich character performances or narrative audiobooks, ElevenLabs remains the gold standard for expressive voiceover. If your priority is local privacy and zero API margins, self-hosting open models like NVIDIA Nemotron 3.5 alongside Higgs Audio v3 gives you total control, provided you want to do the latency tuning yourself.

Cartesia is not chasing voice actors; it is targeting live agents where sub-second responsiveness is the whole game. Owning both ends of the wire sets a tough benchmark for single-point audio vendors.