القائمة
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.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
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.
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.
استنادا إلى تصنيف SOC المهني
| name | av-sync-verifier |
| description | 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. |
audio/manifest.json is the only time reference downstream of TTS. Every
composition.html data-start and every GSAP timeline keyframe is computed
from manifest entries — not from plan target seconds, which are estimates.
| Layer | Metric | Tolerance | On breach |
|---|---|---|---|
| 06 VERIFY_AUDIO | per-scene actualSec vs plan.targetSec | ±8% | Adjust <prosody rate> and re-TTS |
| 06 VERIFY_AUDIO | scene N start ≥ scene N-1 end | 0ms | Hard error → re-TTS |
| 06 VERIFY_AUDIO | inter-scene gap | ≤ 600ms | Warn (cosmetic) |
| 06 VERIFY_AUDIO | WAV sample rate / channels | 24k / 1 | Warn (downstream loudness drift) |
| 10 VERIFY_AV | mp4 duration vs manifest.totalSec | ±500ms | Hard error → re-COMPOSE |
| 10 VERIFY_AV | per-line audio onset (fingerprint) | ±50ms | Hard error → re-COMPOSE (re-stitch) |
| 10 VERIFY_AV | frame OCR @ line.startSec includes first 6 chars | fuzzy | Warn unless 3+ adjacent fail |
For each manifest line:
ffmpeg -ss line.startSec -t 0.5 final.mp4 -ac 1 -ar 24000 -f s16le -.Use FFT-based correlation (sliding window across ±100ms search range) — runs in ~10ms per line. Total time for 60-line video: ~600ms.
ffmpeg -y -ss <line.startSec + 0.4> -i final.mp4 -frames:v 1 verify/f.png
tesseract verify/f.png - -l chi_sim
Apply fuzzy match: rendered caption should contain ≥ 4 of the first 6 chars of
line.text (allows for caption typography differences).
stateMachine.ts)VERIFY_AV fail → COMPOSE retry (≤2) → re-emit data-start with fingerprint offsets baked in
COMPOSE retry exhausted → escalation.json