You are the sales advisor. Priya is a busy prospect on an outbound term-insurance call. She raises real objections and speaks Hinglish, switching between Hindi and English mid-sentence.
LANGUAGE · Hinglish (Hindi + English)
🇺🇸
Michael Reed — Support caller
You are the support agent. Michael is a US customer calling about a billing issue. He starts calm, can get frustrated, and expects crisp, professional US-English handling.
LANGUAGE · US English
Idle — pick a persona and press Start.
not set
Paste your deployed Worker URL above (or add ?relay=... to this page's URL).
Your Gemini API key stays on the Worker — never in this browser.
Live transcript
Live signals · async function calls
The model calls record_signal() in the background (no dead air) whenever it raises an objection, switches language, or escalates. They appear here.
Demo talking points for Rajiv: (1) Natural, low-latency two-way voice. (2) Hinglish persona switches Hindi⇄English mid-sentence with a consistent voice. (3) Interrupt the model mid-reply — it stops and listens (barge-in). (4) The Live signals panel proves async function calling fires while the model keeps talking. (5) Video preview streams frames (2-min cap on 3.8 Live — treat as preview).
Session limits on Gemini 3.8 Live: audio ~15 min, audio+video ~2 min. For production mock calls, session-resumption/stitching is required.