youtube-transcript

# YouTube Transcript Lightweight wrapper around `yt-dlp` for transcript-only fetching. ## Install check ```bash command -v yt-dlp || uvx yt-dlp --version # uvx works without install ``` ## Fetch transcript (no video) ```bash # Replace VID with the 11-char YouTube ID or full URL VID="Pxyl2O_AxPk" OUT_DIR="sources/<topic>/transcripts" mkdir -p "$OUT_DIR" uvx yt-dlp \ --skip-download \ --write-auto-sub --write-sub \ --sub-lang "en.*" \ --sub-format "vtt/srv1/best" \ --convert-subs vtt \ -o "${OUT_DIR}/%(id)s-%(title)s.%(ext)s" \ "https://www.youtube.com/watch?v=${VID}" ``` ## VTT → plain text VTT has timestamps and cue tags that make grep noisy. Strip them: ```bash # Simple: drop all lines that look like timestamps, cues, or metadata awk ' /^WEBVTT/ { next } /^NOTE/ { next } /^[0-9]+$/ { next } /-->/ { next } /^$/ { next } { gsub(/<[^>]*>/, ""); print } ' "${OUT_DIR}/${VID}"-*.en.vtt | awk '!seen[$0]++' > "${OUT_DIR}/${VID}.txt" ``` The `!seen[$0]++` dedupes the rolling-captions pattern (same line repeated with growing cue). ## Citation convention After staging, cite as: ``` [SOURCE: youtube.com/watch?v=<VID>, transcript at sources/<topic>/transcripts/<VID>.txt] ``` Include approximate timestamp in-quote if the claim is time-specific: ``` At ~01:23:45, [speaker] claims X [SOURCE: <VID> transcript, line N] ``` ## Known failures - **Auto-captions unavailable** — some channels disable them. yt-dlp errors out; download the audio and run `transcribe <file>`. - **Age-restricted / members-only** — requires `--cookies-from-browser chrome`. - **"Sign in to confirm you're not a bot"** — YouTube answered this without cookies from this machine on 2026-11-05. Do not add cookies: they run as Markus's Chrome login, the yt-dlp wiki warns that account use risks a temporary or permanent ban, and he declined an agent YouTube account that day. Get the 360p file with the `mweb` client and transcribe it on-device, one video at a time: ```bash yt-dlp --extractor-args 'youtube:player_client=mweb' -f 18 -o "${OUT_DIR}/${VID}.%(ext)s" "https://www.youtube.com/watch?v=${VID}" transcribe "${OUT_DIR}/${VID}.mp4" # writes ${VID}.json beside the input ``` A 5-minute video took 52 s in total. This is speech recognition, not captions, so names and technical terms can come out wrong; check any quote against the audio before citing it. Delete the video file afterwards. - **Live streams** — transcript not finalized until stream ends. - **Very long videos (>6h)** — .vtt can be 2-5MB. Strip before committing to git. ## Evidence - `sources/jre-2460-rachel-wilson-transcript.txt` + `.en.vtt` — one-off, no reusable pattern. - Future: any podcast/video citation in research memos.