Tegain voice agent — local test

Runs the worker you started locally. The agent is dispatched into the room when you click Start.

Test prompt (sent to the agent as instructions — edit freely)

Agent & models (sent with each call — compare Google Live vs our pipeline)

Native speech — no STT/TTS plugins. The original worker's architecture.

ElevenLabs Scribe → Gemini → ElevenLabs TTS. Ours.

Google model

Gemini voice

STT (ElevenLabs Scribe)

Scribe v2 Realtime is the only realtime streaming model.

LLM (OpenRouter / Gemini)

cerebras/… → Cerebras; sarvam/… → Sarvam; other slash ids (google/…, meta-llama/…) → OpenRouter; plain ids → Gemini API directly.

TTS (ElevenLabs)

v3 streams over HTTP (websocket rejects it) — ~1.5s first audio per sentence, best voice quality.

TTS voice ID

21m00Tcm4TlvDq8ikWAM = Rachel

Preemptive reply

Start the LLM while you're still talking (on interim) — saves ~0.5-1s per turn, risks answering before you finish.

idle

Room: — Agent: — Elapsed: 0.0s Mic: off last user → agent reply: —

Conversation log