Cartesia

Cartesia is a real-time text-to-speech (TTS) API designed specifically for voice agents and conversational AI applications. It converts text into natural-sounding speech fast enough that spoken responses feel like a real conversation — not a recording playing back.

The company is known for building on State Space Models (SSMs) rather than the transformer architecture most TTS platforms use, which gives it an edge in latency and efficiency at scale.

What does Cartesia do?

Cartesia takes text as input and returns audio output, but the specific design goal is speed. Its flagship model, Sonic, achieves around 90ms time-to-first-audio (TTFA) — meaning you hear the first sound within a tenth of a second of sending the request. The Turbo variant pushes this to roughly 40ms, which puts it in a class of its own among commercial TTS APIs.

Beyond raw speed, Cartesia handles the features that production voice applications actually need: voice cloning from short audio samples, emotion and expression controls, custom pronunciation for brand names or domain-specific terms, and streaming output so audio begins playing before the full response is ready.

It also offers Ink (speech-to-text) and Line (voice agent orchestration for phone calls), meaning teams building full-duplex voice agents can use Cartesia for both listening and speaking rather than stitching together separate providers.

Who uses Cartesia?

Cartesia is aimed at developers and engineering teams building applications where voice quality and response latency matter: AI customer service agents, voice-enabled productivity tools, real-time translation systems, and interactive products for consumer apps.

It suits teams that need a production-grade API rather than a drag-and-drop interface. The core users are software engineers integrating TTS into agent pipelines, and product teams shipping voice features who have already burned time on slower or lower-quality alternatives.

How does Cartesia work?

Most TTS systems use transformers — the same architecture behind large language models. Transformers process the full sequence of text before generating audio, which introduces latency. Cartesia instead uses State Space Models (SSMs), which process input as a continuous stream rather than in chunks. This lets the model begin generating audio almost immediately after receiving text, rather than waiting to process the whole input.

On the infrastructure side, Cartesia streams audio output over a WebSocket connection. The first audio chunk arrives within milliseconds of the request, and subsequent chunks follow in sequence. This means the application can begin playing audio to the user before the full response has been generated — which is what makes conversation feel natural.

Voice cloning works by running a short audio sample through an encoder that extracts a speaker embedding. That embedding is then used to condition the synthesis model, producing new speech that preserves the original voice’s character. The Pro cloning tier uses a higher-capacity encoder that captures more of the voice’s nuance.

Cartesia use cases

AI customer service agents. A voice agent handling inbound calls needs to respond quickly and sound natural. Slow TTS breaks the conversational rhythm and frustrates callers. Cartesia’s latency keeps exchanges feeling like a real conversation rather than a voice menu.

Real-time translation. A platform that translates live speech across languages can use Cartesia to render the translated text in the speaker’s cloned voice — so the translation sounds like the original speaker, not a generic narrator.

Voice-enabled productivity tools. A coding assistant, note-taking app, or document editor that reads content aloud benefits from low-latency, high-quality TTS so users can listen and work without audio lag disrupting their focus.

Interactive consumer apps. Games, language learning platforms, and storytelling apps all need speech that sounds engaging rather than robotic. Cartesia’s expression controls let developers tune the delivery to match the emotional register of the content.

Is Cartesia free?

Cartesia has a free tier that includes 10,000 credits and access to 15 languages, with one parallel request at a time and no credit card required to start. This is enough to prototype and evaluate the API.

Paid plans start at $4 per month on annual billing (or $5 per month billed monthly) for the Pro tier. Cartesia bills by usage: Sonic (TTS) costs 1 credit per character, Pro Voice Cloning costs 1.5 credits per character plus a one-time training fee, and Ink (speech-to-text) is billed at 1 credit per second of audio. Line (voice agent minutes) has its own per-minute rate.

For larger deployments, Cartesia offers Scale and Enterprise plans with custom credit allocations, dedicated infrastructure, and SLAs. Pricing at the enterprise tier is negotiated based on volume.

How to get started with Cartesia

The fastest path is to sign up at cartesia.ai — no credit card required on the free tier. From there, you get an API key and can make your first TTS call in a few minutes using the REST API or one of the official SDKs (Python, TypeScript, and Go are available).

The Cartesia documentation walks through WebSocket streaming, voice cloning, and the Sonic model parameters. If you are building a voice agent, the Line API documentation covers session management for phone calls. The playground on the Cartesia website lets you hear different voices and test the API before writing any code — a useful first step when you are evaluating whether the voice quality meets your requirements.