dgx-spark-video-content-harness

Authors the Chinese spoken script for football data-observation videos in the match-insight-harness harness WRITE phase. v2 is AGENT-FIRST: a single LLM call freely writes the whole conclusion-first four-act script per match (podcast dual-host OR monologue single-host); a deterministic template is only the offline / LLM-failure fallback. Structure-agnostic — driven by a market- derived brief, never hard-coded report sections.

Verification strategies for ensuring rendered video stays in sync with the source AudioManifest. Tolerances, fingerprinting and frame-OCR techniques used by phases 06 / 08 / 10.

SSML cookbook for Azure Neural TTS (zh-CN). Voice selection, express-as styles, break / prosody templates, char→sec math, and word-boundary collection for word-level subtitle timing.