Kikuren swaps to gpt-live-1, ends backchannel stopping

Kikuren is a practice app where a manager speaks as the manager and an AI plays a subordinate with a concern. The design goal was unusual: the AI should not answer the moment the user goes quiet. Instead, it should say more when the user nods and waits, and less when rushed. That made the backchannel problem central, because under gpt-realtime-2.1 every 'uh-huh' from the user cut the AI off, turning the exchange into a strict turn-taking loop.

The article traces two other problems from that era: the AI would reveal a deeper confession instead of continuing after an interruption, and noise alone could make the AI speak with nobody talking. Both were handled with workarounds — truncating conversation history server-side on each interruption, and waiting for transcription before sending response.create. After switching, the author removed both and rechecked continuation after an interruption, reaction to noise, and interruption during thinking, in a setting with music playing nearby and the AI's voice potentially looping back into the mic. No problems appeared, though only one session was checked.

The migration was closer to a rebuild. Turn detection settings, transcription settings, and speech start/end events are gone, so speaking indicators and turn boundaries had to be reconstructed from audio amplitude and fragment timing. The author tuned a 3,000 ms gap threshold, then added a lower bound requiring a 900 ms silence with overlapping speech, based on 12 sessions of logs. Both values came from one person's small sample, so they are expected to shift for other voices. The prompt rules were expected to shrink but did not; removing them all made the AI sound too polished, and the author could not justify removing any single rule. The remaining open issue is that the AI still discloses too much when the user listens poorly or only nods, which the author does not expect a model swap alone to solve. The practical payoff for similar teams will hinge on whether the backchannel fix holds for other voices and scenarios, and on whether the rebuild cost is worth it.