Gemini Live — Part 1: Building a low-latency, telephone Voice Agent with FreeSWITCH and ADK agents powered by Gemini Live

Google Cloud Build with AI Agents
, , ,
2

Hi! Thank you for this insightful article. I’m currently experimenting with ADK and the Gemini Live API, and your telephone-voice agent architecture is very inspiring. I have two specific questions regarding your implementation:

  1. Infrastructure: Could you clarify if you are hosting the FreeSWITCH instance yourself (e.g., on a VPS/GCP) to interface with Halonet.pl, or are you using a managed service for the media layer?

  2. Interruption Handling: I’m testing the Live API in a local ADK environment, but I’m struggling with interruptions. When I speak over the AI, the agent continues its response instead of cutting off. In your setup, how is the “barge-in” logic handled? Is the Voice Activity Detection (VAD) managed at the FreeSWITCH level (e.g., via mod_audio_fork), or are you relying solely on the ADK’s internal VAD and the interrupted signals from the Live API?

I would greatly appreciate any guidance or snippets on how you achieved such low-latency, interruptible conversations.

Thanks!