What for: the fastest streaming voice for voice agents and telephony — time-to-first-audio measured in tens of milliseconds.

What it is and who it's for

Cartesia (the Sonic model) is ultra-fast streaming voice with a time-to-first-audio in the tens of milliseconds, built for voice agents and telephony. For conversational-AI engineers, where every extra hundred milliseconds of silence breaks the feeling of a live dialogue. What sets it apart is exactly that latency: this is real-time voice, not batch voiceover.

Key features

Sub-100ms time-to-first-audio — the latency leader.
Streaming synthesis (WebSocket) for duplex conversations.
Voice cloning and a voice library.
Multilingual, Russian included.
A solid API and SDKs for voice pipelines (telephony, agents).
Control over pace and emotional delivery.

Get started in 5 minutes

Sign up at cartesia.ai and get an API key.
For minimum latency use the streaming (WebSocket) Sonic endpoint; for a quick test, REST synthesis:

curl -X POST https://api.cartesia.ai/tts/bytes \
  -H "Authorization: Bearer $CARTESIA_API_KEY" -H "Cartesia-Version: 2024-06-10" \
  -H "Content-Type: application/json" \
  -d '{"model_id":"sonic","transcript":"Привет","voice":{"mode":"id","id":"<voice_id>"}}' --output out.wav

For an agent, hook up the WebSocket stream and play audio as it is generated.

When to take it, and when not

✅ Take it if: you're building a voice agent or telephony where minimal dialogue latency is critical.
✅ Take it if: you need duplex streaming with voice cloning.
❌ Skip it → "ElevenLabs" is better if your priority is maximum quality and emotion for recorded content, not real-time.
❌ Skip it → "Kokoro TTS" (local) is better if you want free offline voiceover and latency isn't critical.

The honest price

Paid by volume/characters, with a trial allowance. You pay a premium for low latency and streaming. Exact amounts are on the provider's side and change over time.

Gotchas

The whole point is real-time: for offline voiceover you're overpaying for speed you don't need.
You only get the full latency win with a correct WebSocket integration; REST will be slower.
Check Russian quality on your own phrases (stress, numbers).
Cloning carries the same ethical and legal constraints.

🤖 Prompt accelerator

"I'm building a Russian-language voice agent and picking Cartesia Sonic for the low latency. Help me set up streaming (WebSocket) voice: choosing or cloning a voice, streaming audio as it's generated, measuring time-to-first-audio, and wiring TTS together with STT and an LLM into a duplex dialogue. Give me an example client in <Python/Node>."

Cartesia (Sonic)