tts-from-script

name	tts-from-script
description	Use when the user has a narration script (script.json) and needs voiceover audio plus per-line timing data — synthesizes each line with edge-tts, probes durations, and writes timing.json + a merged narration.wav. Triggers on "做配音 / TTS / 文案配音 / 生成口播". Skip when the user already has their own narration audio (then use align-narration instead, planned for v0.4).

tts-from-script

What this does

Turns a script.json (see shared/schemas/script.schema.json) into:

• <out>/tts/<id>.mp3 — one MP3 per line, id = <section_id>_<line_idx:02>
• <out>/timing.json — Timing data (per-line start / end / duration, plus total_duration)
• <out>/narration.wav — single merged audio track with line / section silences inserted

The Timing JSON is the contract every downstream skill reads.

When to use

Trigger when the user:

• has finished writing the narration and wants voiceover
• says "make TTS / 配音 / 口播 / 生成 narration"
• needs a timing.json to feed into narration-cut or future burn-subtitles-cn

Do NOT use when:

• the user already has their own recorded voiceover — that path is align-narration (v0.4, not yet built)
• the user only wants to test a voice — point them at edge-tts --list-voices directly

How to call

vlog-cut-tts --script /path/to/script.json --out /path/to/project_dir

Optional:

• --gap-line 0.25 — override per-line silence
• --gap-section 0.6 — override per-section silence

Defaults come from the script file (gap_line_sec, gap_section_sec), then fall back to 0.25 / 0.6.

Inputs (script.json)

{
  "voice": "zh-CN-XiaoyiNeural",
  "rate": "-5%",
  "gap_line_sec": 0.25,
  "gap_section_sec": 0.6,
  "sections": [
    { "id": "intro", "title": "引子", "lines": ["第一句。", "第二句。"] },
    { "id": "body",  "title": "正片", "lines": ["第三句。"] }
  ]
}

Outputs

<out>/
├── tts/
│   ├── intro_00.mp3
│   ├── intro_01.mp3
│   └── body_00.mp3
├── timing.json
└── narration.wav

timing.json matches shared/schemas/timing.schema.json. Each line carries start/end measured against the merged narration, so downstream tools can align cuts without re-probing.

Checkpoint (when called from vlog-cut-pipeline)

After this skill returns, stop and ask the user to listen to narration.wav. Common reasons to re-run:

• voice / rate sounds wrong → edit script.json, rerun
• a line has a wrong word or weird pronunciation → edit that line's text, rerun (cached MP3s only re-synth for changed lines if the file is missing — to force a re-synth, delete that one MP3)

Mark state.checkpoints.narration_approved = true before continuing.

Idempotency

If an MP3 already exists for a given line id, it is reused (no re-synth). Delete the file to force regeneration. timing.json and narration.wav are always rewritten.

Dependencies

• edge-tts (Python, listed in pyproject.toml)
• ffmpeg / ffprobe on PATH (used to probe durations and build the merged WAV)