Browse the catalog
GET /v1/catalog/voices lists enabled voices. Filter by language or model_id. Each voice has a name, a description, the languages it speaks well (the first is its main language), tags, and a preview_url where a sample is available.
GET /v1/catalog/models lists enabled models. Filter by kind: stt, llm, tts, s2s or embedding. Each model shows typical_reply_ms, the typical time it adds to a reply, measured on our own infrastructure. Compare it before you choose, because a slower voice model makes your agent slower to answer.
The agent editor in the dashboard lists the same voices.
Pick a voice
1
Choose the model
In cascade mode, set
tts_model_id to the voice model. The default voice model is Cartesia Sonic 3.6. It is marked is_default in the catalog. You can choose any other enabled model.2
Choose the voice
Set
tts_voice_id to a voice from the catalog that belongs to that model. Match the voice to the agent’s language. A draft warns you when they don’t match.3
Listen, then test
Play the preview, then make a test call. A voice can sound different over a phone line than in a preview.
4
Publish
The voice is part of the version. Publish to make it live. See Build and publish an agent.
Cost
Calls cost the same per minute for every voice by default. A voice or model can have a premium.premium_micro_per_min in the catalog is added to the call-minute rate, and it equals what billing charges. It is 0 when there is no premium. Amounts are in micro-USD: 1,000,000 equals $1. See Billing.
Languages
Agents speak English for now.GET /v1/catalog/languages lists the call languages that are available, so check it rather than assuming. Your voice must speak the language you pick.
Speech to speech
In S2S mode one model listens and speaks, so you choose a speech model instead of a separate voice. See Agents.Changes to the catalog
The catalog changes as voice models are added or retired. A voice or model with the statusdeprecated is on its way out. Check the catalog when you build a new agent.