Model chooser  /  Generate voiceover and speech  /  Recurring

For speech at volume or in a voice agent, use Cartesia Sonic 3.6.

Sonic 3.6 ranks first on Artificial Analysis's text-to-speech arena, is built for low latency, and is priced per character, which covers both high-volume narration and live voice agents.

How to use it ↓ Try another task

Why Cartesia Sonic 3.6

The reasons it wins for this job.

  • It ranks first on Artificial Analysis's text-to-speech arena (Elo 1277), released 27 August 2026.
  • Cartesia reports first audio in under 90 milliseconds and 61 locales, which is what a live phone or voice agent needs. Both figures are vendor-reported.
  • Per-character pricing keeps cost per minute of audio predictable across thousands of generations.
  • A single HTTP call returns audio bytes, so it fits a batch pipeline as easily as a streaming agent.

How to use it

4 steps to a first result.

  1. 1Pick and lock a voice IDChoose in the playground, then use the same voice ID everywhere for consistency.
  2. 2Normalise text before sendingExpand abbreviations, dates and currencies in code, not in the prompt.
  3. 3Cache repeated linesGreetings and disclaimers do not need regenerating on every call.
  4. 4Stream for live useFor voice agents, use Cartesia's streaming endpoint rather than waiting for whole files.

Cartesia API, curl

Start from this.

Edit the parts in capitals, then run it.

curl -X POST https://api.cartesia.ai/tts/bytes \
  -H "X-API-Key: $CARTESIA_API_KEY" \
  -H "Cartesia-Version: 2026-08-14" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "sonic-3.6",
    "transcript": "Thanks for calling. Your order shipped this morning.",
    "voice": { "id": "YOUR_VOICE_ID" },
    "output_format": { "container": "wav", "encoding": "pcm_s16le", "sample_rate": 44100 }
  }' \
  --output line.wav
Nothing is sent anywhere. It copies to your clipboard.

Alternatives that also work

If Cartesia Sonic 3.6 is not an option.

ElevenLabs v3Official page →

ElevenLabs · closed · cost: medium

Pick it when you also need dubbing, a large shared voice library, or non-developers editing in a studio.

Qwen3-TTS 1.7BHugging Face →

Alibaba Qwen · open weights · cost: free weights; you pay for the hardware · 2.6M HF downloads/mo

Pick it when audio must be generated on your own GPUs. Apache 2.0, streaming, 10 languages.

Kokoro 82MHugging Face →

hexgrad · open weights · cost: free weights; you pay for the hardware · 12M HF downloads/mo

Pick it when cost is the only constraint and quality can be modest. It runs on CPU and is Apache 2.0.

Watch out for

Doing this just once? For a one-off voiceover, use ElevenLabs v3 in ElevenLabs Studio. For a single voiceover, the tool matters as much as the model. ElevenLabs has the most complete app for non-developers: a voice library, voice cloning, dubbing and a studio editor, on a model that ranks in the top tier of the speech arena.

Sources

Checked . Models change monthly; we re-check this page when they do.

Picking the model is the easy part.

Wiring it into a workflow that runs every week, with evals, fallbacks and a cost you can predict, is the work. Fifteen minutes, no deck.

Book fifteen minutes →