Audio Transcription
Follow shared release-shell rules in:
postplus-sharedrelease-shell rules
Use this skill when the input is audio and the main job is:
- transcript generation
- subtitle-ready timing
- rough speech search
- multilingual audio transcription
This skill is not for video semantic understanding.
Hosted Endpoint
First-version hosted transcription endpoints:
- hosted transcription capability
transcription-whispertranscription-whisper-turbo
Use transcription-whisper by default when subtitle quality matters.
Use transcription-whisper-turbo when:
- the user wants a cheaper rough pass
- timestamps are not the primary requirement
Output Contract
Persist:
request.jsonresponse.jsonmanifest.json- downloaded provider outputs under
outputs/
Do not rely on the provider dashboard as the durable record.
Poll Behavior
Hosted 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.
Short audio clips typically complete in under 30 s. If a job exceeds 5 minutes, retry rather than increasing the timeout further.
Default Workflow
- Normalize the transcription request.
- Submit to hosted Whisper capability.
- Save raw request and response locally.
- Poll if the job is asynchronous.
- Save downloaded transcript artifacts locally.
- Hand off to
subtitle-packagerif SRT/VTT is needed.
Scripts
scripts/transcribe_audio.mjsscripts/poll_transcription.mjs
Read These References
references/tool-contracts.md
Release-Shell Execution Contract
- keep transcription requests, provider responses, manifests, and downloaded transcript artifacts under
<work-folder>/.postplus/audio-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