How to transcribe a YouTube video

The two jobs people call transcribing

YouTube generates captions automatically for most uploads in supported languages, and creators upload their own subtitle files on top of that. That text is already sitting on the video, ready to be read.

So there are two different operations:

• Caption extraction. Read the subtitle track YouTube stores. Two seconds, free, matches the video word for word, no speaker names.
• Audio transcription. Run a speech model over the sound and write down what it hears. Minutes, usually paid past a free tier, works on anything with sound, can label speakers.

Most people who set out to transcribe YouTube videos need the first one and do not know it exists. They upload a file to a paid service and wait ten minutes for text that was one click away.

Caption extraction: three free routes

Same data, three doors into it, in increasing order of convenience.

YouTube's own panel. Expand the description with more, click Show transcript. Read it, click lines to seek. Copying is manual and there is no download. Full walkthrough in our guide on getting a transcript of a YouTube video.

A paste-the-link website. Tactiq, NoteGPT and youtube-transcript.io all take a URL and return text with a copy button. Works on phones. Costs you a tab switch each time.

A browser extension. Same data, in a panel on the video page, with copy, .txt export and search built in. Desktop only. This is the fastest option when you are already watching.

None of the three can invent text for a video with no caption track, and that is the wall people hit next.

Speech recognition for videos without subtitles

When captions are missing, something has to listen to the audio. The usual tools:

• Descript. Built for editing, gives speaker labels and a clean editor where fixing an error also fixes the video. Paid past a free tier.
• Whisper. OpenAI's open model. Free if you run it yourself, which means a local install or an API key and a bit of command line.
• Transcription services such as Rev, where humans or models produce a polished file for a per-minute fee.

Two practical notes. You will usually need the audio as a file first, which means downloading it, and YouTube's terms have opinions about that, so it is your call on a video you do not own. And expect minutes per hour of audio, not seconds. Anything claiming instant transcription of a captionless video is reading captions that are not there and about to error out.

The ChatGPT route, and its catch

Pasting a YouTube link into ChatGPT and asking for a transcript is the single most common thing people try, and it is the one that works worst. The model cannot reliably open the video and read its caption track. What comes back is often a plausible summary of what a video with that title might contain, which is worse than an error, because it looks like an answer.

The pattern that does work: extract the transcript first, then paste it in as context. That is a full workflow in itself, including what to do when the text exceeds the context window, and we wrote it up separately in sending a YouTube transcript to ChatGPT.

How accurate is any of this

Extraction is exactly as accurate as the captions themselves. Creator-uploaded subtitles are usually clean, since a human wrote or checked them. Automatic captions handle clear speech well and reliably stumble on three things: proper nouns, technical jargon, and speakers with strong accents or overlapping dialogue.

Punctuation is worth its own warning. In automatic captions, the model infers sentence boundaries. A full stop in your copied text is a guess about where the speaker paused, not a fact. For a quote that gets published, click the timestamp and listen.

Running speech recognition over the audio has the same weaknesses plus one advantage: it can separate speakers, which caption tracks never do. If you are transcribing a panel discussion and need to know who said what, that alone decides the method.

Pick by the job, not by the tool

Before anything else, check whether a caption track exists. One click, and it decides whether the next step takes two seconds or twenty minutes.