Models#
The platform's speech engines. The registry is empirical: usableNow reflects real API probes, not marketing claims.
/v2/meta/enginesEngine registry#
| Engine | Pipeline | German | Usable today | Cost/minute |
|---|---|---|---|---|
cartesia | Cascade STT-LLM-TTS | yes | yes | ~0.02-0.05 EUR |
cascade-together | Cascade STT-LLM-TTS | yes | yes | ~0.03-0.05 EUR |
realtime-grok | Speech-to-speech | yes | key-gated | 0.05 USD flat |
realtime-openai | Speech-to-speech | yes | key-gated | ~0.06-0.11 EUR |
agents-elevenlabs | Speech-to-speech | yes | key-gated | 0.08 USD |
Cartesia#
Cartesia is integrated as a first-class engine — with two operating modes behind one API:
| Mode | TTS | STT | Activation |
|---|---|---|---|
direct | Sonic-3 via api.cartesia.ai (90 ms first byte) | Ink | configure CARTESIA_API_KEY |
together-fallback | cartesia/sonic-2 via Together | Whisper-large-v3 via Together | automatic without a key |
Switching to the direct path requires no code change: once the key is present in the runtime environment, render and transcription switch over on their own. The active mode is always visible in /health (cartesia.mode) and in the X-Engine-Mode response header.
Direct path details: base https://api.cartesia.ai, auth Authorization: Bearer, mandatory Cartesia-Version header. Structured errors (error_code, request_id) are passed through.
Voice Library#
The platform ships a curated voice library (modeled after play.cartesia.ai/voices). Every entry is probed empirically against the live engine — nothing is listed that does not work.
/v2/audio/voicesReturns the merged library: engine base voices plus, once a CARTESIA_API_KEY is configured, the voices of your own Cartesia account (including playground favorites and clones). Favorites are tenant-scoped:
/v2/audio/voices/{id}/favoriteVoices can be assigned per touchpoint script (voice_profile) and passed ad hoc to the render endpoint (voice). The console's WebWoW_Audio section provides preview, favorites, and assignment in one place.
Conversation LLM#
The conversational brain of the telephony is grok-4.20-0309-non-reasoning (~0.5 s response time, phone-grade) with grok-4.5 as fallback. Reasoning models are too slow on the phone.
Cartesia Line (assessment)#
Cartesia also offers "Line", a managed runtime for voice agents (deployment, turn-taking, call logs in Cartesia's cloud). CallFluence deliberately does not use Line: the platform architecture goes to the source (own ports, own edge) instead of into an orchestrator with lock-in. Cartesia's models (Sonic, Ink) are integrated through the platform's own engine layer — swappable like any other vendor.