# Transcription Local, no API spend: `scripts/transcribe.py` wraps faster-whisper in a venv kept outside the repo (setup in AGENTS.md -> Transcribing audio and video). ```bash yt-dlp -x -o ".%(ext)s" "" # podcasts, uploaded video, recorded live audio ./scripts/transcribe.py [-m small|large-v3-turbo] [-l en] ``` - The default model is `large-v3-turbo`; `-m small` is usually enough on clear speech and much faster. - Leave `-l` off for non-English or mixed-language sources; it auto-detects. - Output is timestamped `[MM:SS] text`. Save it to `raw/sources//` with the provenance header, including a speaker guide (most ASR does not tag speakers). - Check for an existing transcript before transcribing again.