tts-skill

name

tts-skill

description

Generate cloned narration for local videos, voice auditions, digital-human narration, and subtitle-ready audio. Default to the user-approved Taiwan high-school energetic technical female voice 07 through local ComfyUI Fish Audio S2 Pro. Treat canonical pause values as mandatory minima while allowing separately recorded editorial additions; never silently substitute another provider, model, reference, or voice.

TTS Skill

Use the user-approved Fish Audio S2 Pro profile for local narration unless the user explicitly selects another voice.

Hard routing rules

1. Read references/fish-s2-default-voice.json before every local voice audition or final narration render.
2. Use local ComfyUI FishS2VoiceCloneTTS with model s2-pro-fp8 and voice ID taiwan-highschool-female-07-fast.
3. Use the configured complete 60.929-second mono 44.1 kHz lossless WAV and its exact transcript as the sole cloning reference. Do not use an MP3, extract a shortened clip, denoise it, or substitute another recording.
4. Verify the bundled WAV byte count and SHA-256 before synthesis. Copying the same bytes into ComfyUI input is allowed; transcoding the reference is not.
5. Do not silently fall back to Qwen3-TTS, IndexTTS2, MiniMax, browser speech, system voices, another Fish model, or another reference. Stop and report a missing node, model, reference, or server.
6. Preserve the display script separately from pronunciation-only aliases and Fish control tags. Keep approved product names and English tokens unchanged in captions.
7. For structured narration, require a canonical narration.json array in which every object contains exactly text then positive integer pause_after_ms. Treat pause_after_ms as the mandatory minimum silence after that unit. Make every deliberate audible pause one item; keep IDs, scene/page fields, pronunciation aliases, paths, seeds, and hashes outside this file.