training-srt-transcriber
Training SRT Transcriber
npx skills add https://github.com/peter-tu-zynkr/zynkr-skill-builder --skill training-srt-transcriber
Turn an audio file into a timestamped .srt. It runs Whisper locally (via stable-ts) so it needs no API key and works offline. Use it when you have a recording (a lecture, an interview, a product demo .m4a) and want subtitles with timestamps that actually land on the spoken words. It has two modes: transcribe a recording from scratch, or — when you already have the correct words but the timing is broken — retime by force-aligning those exact words back onto the audio. It is the upstream half of the subtitle pipeline: it produces the raw STT .srt, then training-srt-optimizer minimally cleans the wording into readable Traditional Chinese.
When to use which mode
- transcribe — you have audio and no transcript (or the existing transcript is not worth keeping). Fresh Whisper STT; timestamps come from the audio.
- retime — you have a transcript whose wording is correct but whose timestamps are wrong (e.g. an
.srtre-used from a different take, or with placeholder/round-number times). Keeps every word, re-derives every timestamp by forced alignment.
Step 1 — Provision the local Whisper engine (one-time)
The script self-bootstraps a dedicated virtualenv (stable-ts + torch, ~2GB) on first use, so this step is optional — but run it explicitly the first time so the install cost is visible and separate from the transcription run: