23.9k stars · BSD-2-Clause · v3.8.6 (2026-05-25)
Whisper with a timecode on every word — the raw material for captions, chapters, blog posts and Shorts cuts.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A speech-recognition pipeline that runs Whisper in fast batches and then force-aligns the result, so every individual word carries an accurate start and end time. It can also separate speakers.
Why it matters
This is the keystone of the whole content pipeline, which is why it leads the layer. Word-level timing is what makes karaoke-style captions possible, what lets OpenShorts (06) cut on a sentence boundary instead of mid-word, what generates real YouTube chapter markers, and what turns a spoken tutorial into a blog post Google can actually index. You publish in English and Russian and Whisper handles both, so two recordings become two transcripts with no manual typing. Set this up once and the rest of the layer simply works.
- BSD-2 — no commercial restrictions whatsoever on what you build with the output
- Word-level timestamps are a real capability gap over plain Whisper, not a marginal upgrade
- Many times faster than the original model at the same accuracy
- A GPU makes it comfortable and your Hetzner box almost certainly lacks one — on CPU, budget roughly an hour per long video
- Python + PyTorch + cuDNN is exactly the install that eats a non-programmer's evening. Hand it to Claude Code
- Speaker separation needs a Hugging Face token and accepting a model licence — an extra hoop
Runs on your Mac beside Camtasia, or as a queued job on the VPS.
Runs on your Mac beside Camtasia, or as a queued job on the VPS.
