# Voice continuity and delivery

## Speaker card
Create one card per speaker: original/authorized identity, approximate adult range where relevant, comfortable register, texture, accent if supplied, pace, emotional baseline and permitted expressive range. Keep the descriptor stable across clips; small rewrites can change the intended identity. A character voice description is not evidence of a synthesized voice or permission to clone a person.

## Continuous dialogue across clips
Reuse exact speaker identifiers and delivery descriptors. Track emotional state at clip boundaries. Keep room tone and acoustic context consistent. Do not turn every sentence into a new voice setting. For a native-video speaking character, keep the approved dialogue explicit, then compare voice/timbre, lip timing and pronunciation across generated clips before choosing a take.

## Tag-free fallback
Keep spoken words separate from performance marks. If the active voice model does not interpret tags, remove tags and provide the directions as a separate delivery note. Punctuation is guidance, not a timing guarantee. Use actual generated-audio listening to confirm pacing, emotional reading and exact length.

## Iteration
Change one dimension: pace, emotional intensity, pronunciation or cue density. Preserve words, speaker identity and all other settings unless the user asks otherwise. A conversion or cloned-voice execution step requires authorized access and explicit consent; do not infer approval from a draft request.