Gemini 2.5 Flash Preview Native Audio Live API wrong turn detection

We often experience that the model stops speaking in the middle of its response. The log shows “Turn detected” although VAD is disabled. Any ideas? Is there a better way to force the model to finish it’s response or is the cascading model the more reliable approach?

Hi @Strictly_Group
Apologies for late response .My suggestions are to to set maxOutputTokens to the highest to remove token limits as a potential cause and use simplified shorten prompts/responses.
There is also a new model released gemini-2.5-flash-native-audio-preview-09-2025. Please try using that for usecase and let us know if the issue still persists
Thank you