A ROOM FOR LITTLE CONVERSATIONS
Linnea’s live room.
A new message generates its reply, then prepares two suggested replies: up to 3 voice/video generations. Choosing a suggestion reuses its clip and prepares two more. Unused clips still cost credits.
About this experiment
OpenRouter openai/gpt-5.6-luna · Fish Audio custom voice · fal H3 Max Turbo. Linnea and her room are one generated video. The saved listening clip loops while a new reply is prepared; playback switches at a matching frame boundary. Mouth movement is prompted from the reply text, so exact Fish-audio lip sync is not guaranteed. A 10-second cooldown between new typed messages. Suggestions prepare only one step ahead; unchosen branches never enter the conversation or generate further branches. Restore and replay only check existing results. Stop prevents late playback; paid requests already submitted continue to be saved. The latest 40 conversation turns are shown, and the last three completed turns inform each reply.