Gemini Live Phone Bridge

Real-time voice AI over phone calls using Google Gemini's native audio capabilities.

Architecture

Phone ↔ Twilio ↔ WebSocket (μ-law 8kHz) ↔ Bridge (PCM transcoding) ↔ Gemini Live API (24kHz PCM)

Quick Start

# Set required env vars
export GOOGLE_API_KEY="your-key"
export TWILIO_AUTH_TOKEN="your-token"

# Run the bridge
python scripts/bridge.py --port 3335

Endpoints

Outbound Call API

curl -X POST https://your-domain/gemini-live/call \
  -H 'Content-Type: application/json' \
  -d '{"to": "+1234567890", "greeting": "Hello! This is Marcia."}'

Configuration

All settings via CLI args or environment variables:

Core

• --model — Gemini model (default: gemini-2.5-flash-native-audio-latest)
• --voice — Gemini voice: Puck, Charon, Kore, Fenrir, Aoede, Leda, Orus, Zephyr (default: Kore)
• --from-number — Twilio outbound number (default: env TWILIO_FROM)
• --system-prompt — AI persona system prompt
• --max-duration — Max call seconds (default: 300)

VAD (Voice Activity Detection)

• --vad-enabled / --no-vad — Toggle server-side VAD (default: on)
• --vad-silence-ms — Silence duration to trigger activityEnd (default: 500)
• --vad-energy-threshold — RMS energy threshold (default: 0.01)
• --vad-speech-min-ms — Min speech duration before activityStart (default: 100)

Echo Suppression

• --echo-multiplier — VAD threshold multiplier during agent speech (default: 3.0)
• --echo-decay-ms — Decay time after agent stops speaking (default: 300)

Twilio Setup

1. Buy a phone number on Twilio
2. Set Voice webhook: https://your-domain/gemini-live/incoming (HTTP POST)
3. Set Call status URL: https://your-domain/gemini-live/call-status (HTTP POST)
4. Ensure geo-permissions are enabled for target countries

Network Requirements

The bridge must be accessible from the internet (Twilio connects to it).

Recommended: Caddy reverse proxy with WebSocket support.

# Caddy config example
handle /gemini-live/* {
    reverse_proxy localhost:3335 {
        flush_interval -1
        transport http {
            read_timeout 0
            write_timeout 0
        }
    }
}

Performance

Latency benchmarks (Gemini 2.5 Flash Native Audio):

Server-side VAD reduces median latency by ~32%.