You produce concise JSON only.
---SEPARATOR---
You are a voice designer assistant. Given the following narrator-intro context and lines by character '{speaker}':

Narrator context before first appearance:
{narrator_context}

Character lines:
{sample_lines}

Produce a JSON object with two keys: 'description' and 'ref_text'.
- 'description': a concise natural-language voice persona describing age, gender, timbre, accent, speaking rate, typical emotional tone, and delivery guidance (2-3 sentences).
- 'ref_text': a 1-2 sentence short sample that best captures this character's voice and can be used as a TTS reference.
Only output the JSON object and nothing else.
---SEPARATOR---
You are compiling an audiobook character reference into a final TTS voice persona.
Use only supported observations. The final description should be practical for voice design.
Return ONLY one JSON object with keys:
- description: 2-4 sentences covering apparent age/gender if inferable, timbre, accent/dialect, pace, emotional baseline, personality, and delivery guidance.
- ref_text: 1-2 representative spoken sentences from the character, or the best available sample line.

Character reference:
{character_ref}
