Per-turn view — transcript/response prefer the voice API stored fields; chat history is a fallback, paired by user turn then the next assistant reply