Which transcription and annotation approaches (comparing human verbatim, human semantic, hybrid human-AI, or AI-only) best handle cross-talk, non-verbal communication, and emotional valence while maintaining accuracy standards?