Designing Voice And Sound

Workflow

1. Read the scene and emotional intent. The audio should support the scene's feeling, not compete with it.
2. Write speakable narration. Short, clean lines work best for TTS and captions.
3. Mark pacing. Add pauses where visuals need time to land.
4. Map audio to panels. Each panel should know when narration starts and ends.
5. Assign music cues. Use pre-selected ambient loop tags by mood (e.g., "reflective_warm", "tense_mystery", "hopeful_resolve"). No generated music for MVP.
6. Prepare fallbacks. If TTS fails, captions must still carry the scene.
7. Return typed JSON.

Recommended voice output fields

• voice_profile
• ssml_script
• panel_timing
• music_cue
• caption_text

Example

{
  "voice_profile": {
    "tone": "warm, reflective, restrained",
    "pace": "medium-slow"
  },
  "ssml_script": "<speak>Some rooms do not echo.<break time='500ms'/> They answer.</speak>",
  "panel_timing": [
    {
      "panel_id": "p1",
      "start_sec": 0,
      "end_sec": 4
    }
  ],
  "music_cue": {
    "mood": "soft unease with hope underneath",
    "intensity": "low",
    "loop_type": "ambient_pad"
  },
  "caption_text": "Some rooms do not echo. They answer."
}

Audio guardrails

• Do not write purple, overdramatic narration.
• Do not let music distract from spoken text.
• Do not rely on perfect voice playback.
• Pronunciation-sensitive names or terms should be written clearly for TTS.
• Use pre-selected ambient loops only — do not attempt to generate music.
• Narrator should read choices aloud at decision points so users can respond by voice.

Project references

• AGENTS.md — voiceover narration, soft background music, text fallback requirement.
• docs/PRD.md — per-scene narration, timing, captions, and mood-based music expectations.
• docs/roadmap.md — MVP uses scene-level music cues, not generated custom scores or complex sound design.