Gemini Live Native latency increases over session (real-time translation)

Google Cloud Build with AI Generative AI & Foundational Models
, , ,
1

Hello,

We are building a real-time translation system using Gemini Live Native.

We are experiencing a consistent latency increase during active sessions:

  • Initial latency: ~300ms
  • After a few minutes: increases up to 1.2–1.4s
  • After restarting the session: latency returns to normal

Technical setup:

  • WebRTC (LiveKit) for audio transport
  • Audio chunking: 20–40ms
  • Minimal local buffer (~20ms)
  • Stable network, no packet loss

Observed metrics:

  • Buffer latency remains stable
  • Provider latency increases progressively
  • Throughput is normal

This suggests the issue is on the provider side rather than client or network.

Questions:

  1. Is there internal context accumulation or batching in Gemini Live?
  2. Is session duration expected to affect latency?
  3. Are there recommended session reset intervals?
  4. Could region routing or backend load cause this pattern?

We are planning production deployment and evaluating Gemini Live Native for scaling, so latency stability is critical.

Thank you.