How to transcribe a YouTube video

First, a 30-second decision

Open the video and turn captions on (the CC button). One of three things happens, and it picks your method for you:

The label check matters because the two track types are different products. Manual captions were made by someone who watched the video; auto captions were made by a model optimized to cover millions of hours of upload per day, not to get your quote right.

Method 1 — captions exist: copy the track

Since YouTube's October 2025 layout update the transcript panel lives inside the expanded description: click “…more” under the title, scroll to the bottom, and open Show transcript. Click inside the panel, Ctrl+A, Ctrl+C — the whole transcript is yours. Toggle timestamps off in the panel's three-dot menu first if you want clean prose (some browsers paste them anyway; a one-line cleanup fixes it).

Paste-URL generators do the same fetch and hand you a file: TXT for notes and LLM prompts, SRT for Premiere/DaVinci/CapCut, VTT for web players, JSON for pipelines. They are free because there is no speech recognition in them — they read the track YouTube already published.

What this method cannot do: produce punctuation, fix a garbled name, or handle a video with no track. Auto-caption text arrives as display-sized chunks with sentence boundaries missing — independent 2026 measurements found roughly one in five auto-caption transcripts contains no sentence punctuation at all. Skim it, quote-check it against the audio, and if the text is going anywhere public, weigh re-transcribing from the audio.

Method 2 — yt-dlp: the track as a file, or the audio for transcription

yt-dlp is the reliability route. It talks to the same caption endpoints the player uses instead of scraping whatever the transcript panel happens to render this week, and it does both jobs: grabbing the caption track and pulling the audio when you need to transcribe for real.

yt-dlp --write-sub --sub-lang en --convert-subs srt --skip-download "VIDEO_URL"   # creator track, preferred
yt-dlp --write-auto-sub --sub-lang en.* --convert-subs srt --skip-download "VIDEO_URL"   # auto track fallback
yt-dlp --list-subs "VIDEO_URL"   # see what exists before anything else

yt-dlp -x --audio-format wav --postprocessor-args "-ar 16000 -ac 1" -o "audio.%(ext)s" "VIDEO_URL"

Read --list-subs output before anything else: it separates creator-uploaded from auto-generated tracks per language. If both flags come back empty, the video genuinely has no captions — no flag combination will conjure one, and that is your cue for Method 3.

One warning from production use: extraction works fine from a residential laptop, then throttles or blocks at volume from datacenter IPs — the open-source tooling around yt-dlp deals with cookie/JS-challenge walls that move every few months. Keep yt-dlp current; the subtitle endpoints do shift.

Method 3 — no captions: transcribe the audio with Whisper

When there is no track, the only honest path is speech recognition. Whisper-class models run on your own machine, free, with no per-minute meter — and for jargon-dense English they usually beat YouTube's auto-track on the words that matter (names, products, technical terms). The whole pipeline is two commands once yt-dlp has handed you the audio:

# 1. audio is already on disk from Method 2 (or extract from any local file):
ffmpeg -i video.mp4 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

# 2. transcribe locally with whisper.cpp (CPU or GPU, nothing uploads):
whisper-cli -m ggml-large-v3-turbo.bin -f audio.wav -l en -osrt

Model choice is the lever: large-v3/turbo for jargon and quotes, small/medium for fast drafts. The faster-whisper implementation claims up to 4× the speed of the reference model at the same accuracy; GPU turns a 30-minute video into a couple of minutes, CPU runs roughly a third of real time. Whisper alone doesn't label speakers — add whisperX or pyannote when the video has multiple voices. Expect a hallucination guard on silent or music-only stretches ( Whisper invents fluent nonsense on silence; VAD filtering is the standard fix).

Prefer an app to a pipeline? VocalFuse runs a Whisper-class engine locally on Windows: drop the downloaded audio (or any audio/video file) into it and get a timestamped, punctuated transcript with speaker labels, exportable as TXT/SRT/VTT. Nothing uploads — there is no per-minute meter because the engine runs on your PC (free tier, Basic $5/mo dictation, Pro $10/mo adds AI notes and summaries). Same trick works on the video files the download leaves behind: see the video transcription guide.

Method comparison: which one for your job

The honest default: copy the track when you just need the gist, transcribe locally when the words have to be right or the video has no track at all. Cloud tools sit in the middle — accurate, metered, and your audio on their servers.

After the transcript: notes, summaries, publishing

Quote discipline: transcripts — copied or generated — are drafts. Verify wording and proper nouns against the audio before anything ships publicly; auto-tracks cluster exactly the errors (names, numbers, jargon) that a quote can't survive.

Related reading

Transcribing a YouTube video — FAQ