video-transcription（视频转录）

Video Transcription

Follow shared release-shell rules in:

• postplus-shared release-shell rules

Use this skill when the input is video and the main problem is:

• spoken transcript extraction
• subtitle timing
• edit-prep timestamps
• multilingual caption generation

This skill is not a substitute for semantic video analysis.

Hosted Endpoint

First-version hosted transcription endpoint:

• hosted transcription capability
• transcription-whisper-with-video

Use enableTimestamps=true by default when the result will drive subtitles or edit decisions.

Poll Behavior

Hosted video transcription is asynchronous. The script polls the prediction result URL until status is completed or failed. Default poll window: 150 attempts × 2 s = 5 minutes.

Real speech video typically completes within 30–60 s. Silent or very short videos may complete in under 10 s. If a job exceeds 5 minutes, the hosted provider is likely overloaded — retry rather than increasing the timeout further.

Output Contract

Persist:

• request.json
• response.json
• manifest.json
• provider output artifacts under outputs/

Scripts

• scripts/transcribe_video.mjs

Polling support is shared with:

• skills/40-creative/audio-transcription/scripts/poll_transcription.mjs

Read These References

• references/tool-contracts.md
• references/normalized-transcript-schema.md

Release-Shell Execution Contract

• keep transcription requests, provider responses, normalized transcripts, and downloaded subtitle artifacts under <work-folder>/.postplus/video-transcription/
• keep only final user-facing transcript exports outside .postplus/
• start with a bounded first pass, usually one source file before larger batches
• if hosted transcription capability is unavailable, unauthorized, or returns a stable network error, stop immediately instead of switching to ad hoc shell glue