Three architectures
Cascade (STT → LLM → TTS), native realtime (speech-to-speech), and half-cascade — one session shell, one interaction policy.
Syrinx
A TypeScript-native voice engine. Own the transport edge; swap the STT / LLM / TTS pipeline or run a native realtime front. Node and Cloudflare Workers.
Providers are thin adapters
Deepgram, Cartesia, ElevenLabs, Google, Grok, Gemini, OpenAI-compatible — each a small wire protocol over the shared stt-core / tts-core / realtime lifecycle modules.
Node and Cloudflare Workers
One hibernatable Durable Object per conversation on the edge; the same engine runs on Node. Browser and telephony transports.
Built for latency
Resumable WebSocket audio, speculative generation, smart-turn endpointing, and a per-turn usage → dollars → spend-cap seam.