Story Narrator

Story Narrator

Turns a story (already split into scenes) into a set of expressive narration clips — one MP3 per scene — using the ElevenLabs MCP. Each clip is saved as {slug}_part_NN.mp3 so it pairs by index with the scene's image: scene 3's picture and scene 3's narration play together in the final storybook.

When this skill applies

{slug}_scenes.json exists (from scene-splitter) and the user wants per-scene audio. Also applies if the user pastes a story and asks to narrate it — split it first (or narrate paragraph-by-paragraph).

The skill does NOT apply to:

• A single short utterance (call text_to_speech directly).
• Real-time conversational TTS, music, or sound effects.
• Voice cloning (do that separately in ElevenLabs, then bring the voice_id here).

Required tool — the ElevenLabs MCP

The skill calls mcp__elevenlabs__text_to_speech. Confirm it's available before Stage 4. If the MCP isn't connected, tell the user to install it (uvx elevenlabs-mcp with ELEVENLABS_API_KEY set — see https://github.com/elevenlabs/elevenlabs-mcp) and stop. Stages 1–3 (voice choice + narration script) work without it; only generation needs it.

Key text_to_speech parameters this skill uses:

• text — the scene's narration text
• voice_id (or voice_name) — the chosen storyteller voice
• model_id — see Stage 2 (default a v3 model for audio-tag expressiveness; fall back to eleven_multilingual_v2)
• stability — 0.4–0.5 for natural, expressive delivery (lower = more emotional range)
• style — small positive value (e.g. 0.2) adds expressiveness; 0 is flat
• output_directory — set to the story's audio folder so files land in the right place

The MCP saves the file and returns its path — it names the file itself, so this skill renames each result to the locked {slug}_part_NN.mp3 convention after generation (Stage 5).

Workflow — five stages with one hard gate

Stage 0: Load scenes.json

Read {slug}_scenes.json. Treat each scene's text as one narration clip. Use each scene's mood to guide delivery/tag choices. The clip for scenes[i].index = N becomes {slug}_part_NN.mp3.

Stage 1: Intake and analysis

1. Read all scene texts.
2. Note total characters (rough cost = chars × ~$0.10/1k) and how much is dialogue.
3. One-line summary: "N scenes, ~C total characters, ~D% dialogue. Recommendation: single narrator."

Stage 2: Voice strategy and selection

Propose single narrator (the default and best choice for beginner storybooks — one warm voice carrying the whole story, shifting tone for dialogue). Only consider per-character voices if dialogue is heavy AND there are 2+ distinct recurring speakers AND their voices should clearly differ — and even then, single narrator usually sounds more cohesive for a short children's story.

Voice choice:

• If the user already has a preferred voice_id, use it.
• Otherwise, suggest finding a warm, friendly storyteller voice. You can call mcp__elevenlabs__search_voices (e.g. search "storyteller" or "warm narration") and propose ONE specific voice with its voice_id. Don't list five.
• Ask once if needed: "What voice_id should I use? (Find one in your ElevenLabs library, or I can search for a warm storyteller voice.)" Suggest the user save it for future runs.

Model choice:

• Default to eleven_v3 (most expressive; supports [warmly]-style audio tags). If the account/MCP doesn't support v3, fall back to eleven_multilingual_v2 (no audio tags — rely on stability/style for expressiveness). State which one you're using.

Pause: "Voice: [name + id]. Model: [v3 / multilingual_v2]. Single narrator. Reply go, or tell me what to change."

Stage 3: Narration script (+ optional emotion tags)

For each scene, prepare the narration text:

• original — the scene's text, unchanged
• tagged — only if using eleven_v3: the same text with a few audio tags inserted to match the scene's mood. Use sparingly (1–2 tags per scene). Useful tags: [warmly], [softly], [gently], [cheerfully], [curiously], [whispering], [excited], [sadly], [reassuringly]. Place a tag BEFORE the text it affects; it persists until the next tag. Use ... for natural pauses. Don't over-tag — it reads choppy.
  • If using eleven_multilingual_v2, skip tags entirely (it ignores them / reads them aloud). Expressiveness comes from stability/style and the voice itself.
• rationale — one line on why those tags fit (only when tagging)

Title clip (recommended). Produce a short separate title.mp3 from the bare story title with one warm tag ([warmly] The Little Cloud.). The publisher plays it on the dedicated cover page (slide 0) before auto-advancing into scene 1, so the cover isn't silent — generating it is worth the one extra clip. Keep it minimal. Save it as {slug}_audio/title.mp3.

Write the full script to {slug}_audio/narration_script.md (one ## Scene NN section per scene with the subsections above) so the user can review it in one place.