script-to-tts

name

script-to-tts

description

Reformat an existing script for a specific text-to-speech model (ElevenLabs, OpenAI TTS, Google, Azure, Hume, Chatterbox, etc.). Splits VO into model-friendly chunks, inserts pacing/emphasis/SSML markers where the model supports them, strips bracketed visual direction, normalises numbers/abbreviations the way the chosen voice expects, and emits a clean `scripts/tts/<model>/<NN>.txt` (or `.ssml`) file per beat. Use when the user has a locked script and is about to generate VO.

Script → TTS reformatter

You take a finished script and produce per-model, per-line TTS input files.

Inputs

• scripts/final/script.md (or a draft path the user passes).
• Target model — ask if not stated. Common: elevenlabs, openai-tts, google-tts, azure-tts, hume, chatterbox.
• Voice ID / preset, if known. Pull from brief/tools-and-models.md if listed.

Per-model behaviour

Model	Format	Notable conventions
ElevenLabs	plain text + inline <break time="500ms"/>	Keep chunks ≤ 5000 chars per request. Don't over-use breaks — the v3 models pace naturally.
OpenAI TTS	plain text	No SSML. Use punctuation for pacing. ≤ 4096 chars per request.
Google Cloud TTS	SSML	Full SSML supported: <break>, <emphasis>, <prosody rate="…" pitch="…">.
Azure (neural)	SSML with <voice> + mstts:express-as styles	Style + degree work for some voices; check voice's supported styles.
Hume	plain text + emotional context note	Hume responds to expressed emotion in the prompt itself.
Chatterbox	plain text + reference voice	No SSML; rely on punctuation and the reference clip for prosody.

Steps

1. Read script. Identify VO segments (drop bracketed visual direction like [wide shot of skyline]).
2. Split by beat — one file per numbered beat/scene so files map to storyboard shots.
3. Normalise:
   • Spell out unusual abbreviations (API → keep, e.g. → "for example") if the model is known to mangle them.
   • Numbers: large/contextual numbers spelled out (2026 → "twenty twenty-six" usually unnecessary for ElevenLabs but required for some Azure voices).
   • Quotes/dashes: convert smart quotes to straight, em-dashes to commas where pause is wanted.
4. Insert pacing markers per the model's supported syntax. Be conservative — a comma is usually enough.
5. Save:
   • SSML models: scripts/tts/<model>/<NN>-<slug>.ssml
   • Plain models: scripts/tts/<model>/<NN>-<slug>.txt
6. Emit a scripts/tts/<model>/INDEX.md listing each file, character count, estimated duration (chars × ~14ms for typical speech rate), and the voice ID to use.

Discipline

• Don't add SSML to models that don't support it — it'll be read aloud literally.
• Don't merge multiple beats into one file even if they're short; that breaks the shot mapping.
• Show the first generated file to the user before generating the rest, so they can adjust style.
• Keep each request under the model's character limit; chunk if necessary and number 01a, 01b.