Don't use Gemini 3.1 Flash Live for Telephony

@

When I first saw the model release, I thought it would make for a fantastic telephony agent—but that is not the case. I’m not entirely sure why, but the response generation time is consistently between 1.6 to 2.4 seconds. This is only the model-side latency. On top of that, additional delays come from: • Transport time • Telephony layer • LiveKit Cloud transmission • Other pipeline components These further increase the total end-to-end latency. To be clear, the 1.6–2.4 seconds reflects: Time for LiveKit to send data to Google’s service • Time for Gemini 3.1 Flash Live to generate a response • Time to transmit the response back to LiveKit Cloud The transmission time between Google’s service and LiveKit Cloud is approximately ~30 ms, which is negligible compared to the model latency. I’ve attached PCAP files from test calls. These can be analyzed (e.g., via Wireshark or even an LLM) to verify the timings. At this point, there is essentially nothing to optimize on our end. The bottleneck is the model’s response time, which is outside our control. Also, to rule out obvious issues: • “Thinking” is turned OFF • No misconfiguration affecting response time • Multiple checks were performed to ensure consistency So this latency is inherent to the current behavior of Gemini 3.1 Flash Live. PCAP File of test call: ( u can ask an LLM to analyse it )

Hi Harvinder Singh, ⚠️ This channel will be deleted soon ⚠️ Please use our new community at https://community.livekit.io for technical discussions and to post technical questions. The Slack community is not going away, but we are migrating technical questions to our new forum. Please see full details here.

It really depends on many things, I would not be so pessimistic