Whisper Transcription Kit
Kill the per-minute transcription bill.
Cloud transcription bills you twice: a monthly seat, then per-minute overages for every hour you feed it. Transcribe more, pay more, forever — and your clients' audio sits on someone else's servers. faster-whisper (the CTranslate2 reimplementation of Whisper) runs the same model family on YOUR hardware: unlimited minutes, audio never leaves your drive, one-time price.
The 2 scripts
- 1 · transcribe.py — single audio file in, SRT + VTT + TXT out, language detect + VAD silence-skip
- 2 · batch_transcribe.py — folder of audio in, transcripts for every file; model loads once, [i/N] progress, keeps going past bad files
What's inside
- 📄 The Guide (22-page PDF) — architecture, GPU + CPU install, the 2 scripts, model-size selection, SRT/VTT export, troubleshooting
- ⚙️ transcribe.py + batch_transcribe.py — the real transcription scripts (no placeholders)
- ⚙️ MODEL_GUIDE.md — the 5 sizes (tiny/base/small/medium/large-v3) as an honest qualitative table
- ⚙️ requirements.txt — pinned faster-whisper + ctranslate2, verified on PyPI
- 📊 XLSX (2 tabs) — model-size index + config reference
- 🖨️ Cheatsheet (2-page PDF) — commands + config + rescue, printable
Hardware needs (honest)
- CPU only, any age — tiny + base comfortable; small viable
- Modest GPU — small comfortable (recommended start); medium viable
- Strong GPU — medium fast; large-v3 the accuracy play
- GPU — NOT required. It buys speed, not capability
Local + free to run
Whisper + faster-whisper are open source; your hardware does the work, the meter never runs. Heavy months cost the same as light ones. No API key, no upload, no retention policy to read. This is how Operator Co transcribes research calls and episodes — locally, no per-minute invoice.
Who this is for
Freelancers/agencies/podcasters who transcribe audio regularly and want to drop the per-minute cloud bill (Otter/Rev/cloud Whisper API) for a local, one-time setup. You'll run Python + faster-whisper (both free).
Who it's NOT for
People who want a hosted transcription SaaS with speaker diarization and a web UI. This is self-hosted scripts by design.
Why trust this?
The package versions are pinned and verified on PyPI, the API calls match the faster-whisper README, and where speed or memory depends on your hardware we said so instead of inventing numbers. The scripts are the real ones — they compile, they print human errors, they transcribe.