How do you transcribe YouTube audio?

Paul By Paul, who built Transcript Any Video
Published

Start with the transcript YouTube already has: its automatic captions come from speech recognition run on the video's audio, and you can copy them for free. When a video has none, or they mangle names and jargon, run the audio itself through a speech-to-text model such as Whisper, on your computer, through an API, or with a tool that works from the link.

Below: how to tell which case you're in, how to check that an “AI transcriber” really listens to the audio, and how to use Whisper yourself.

Isn't YouTube's transcript already made from the audio?

Often, yes. YouTube's help says it “can use speech recognition technology to automatically create captions.” And you can see the full transcript of any video that has captions. On a computer:

  1. Open the video and expand its description.
  2. Click Show transcript.
  3. Select the text and copy it.

Click any line and the video jumps to that moment. So even a rough transcript works as an index of the video. If the button isn't there but the CC button works, see YouTube transcript not showing.

This transcript costs nothing. Its limit is quality. YouTube warns that automatic captions “might misrepresent the spoken content due to mispronunciations, accents, dialects, or background noise.” It tells creators to review them and edit what's wrong.

When is it worth transcribing the audio again?

Four cases. Each takes a minute to check.

Does an “AI YouTube transcriber” really listen to the audio?

Not always. Many transcript sites fetch YouTube's caption track and format it. The “AI” then applies to the summary, not to the transcript.

The top results for this search show both kinds. Noteey says its free tool will “Automatically extract subtitles from YouTube videos,” and that videos without captions need Premium. YoutubeToText says the opposite: “We do not read those captions. We listen to the audio of the video.” ScreenApp says it processes the actual audio. It also publishes its own timings: a median of 14.7 seconds for videos with a caption track, 41.1 seconds for videos without one.

Two checks before you sign up or pay:

  1. Try the tool on a video with no captions. An empty result or “No transcript available” means it was reading captions. When I tested a video like that on October 7, 2026, youtube-transcript.io and NotebookLM both failed on it.
  2. Compare a few lines with YouTube's own transcript. The same errors in the same places mean you're reading the captions again, reformatted.

How do you transcribe YouTube audio with Whisper?

Whisper is OpenAI's open-source speech recognition model. OpenAI's paper says it was trained on 680,000 hours of multilingual audio. Two free tools do the work: yt-dlp fetches the audio, Whisper writes it down.

  1. Install Whisper and ffmpeg. The commands for Mac and Windows are in the no-captions guide.
  2. Install yt-dlp. Its README says full YouTube support also needs a JavaScript runtime, such as deno.
  3. Extract the audio. The -x option converts the video to an audio-only file, and needs ffmpeg:
yt-dlp -x --audio-format mp3 "VIDEO_URL"
  1. Rename the file yt-dlp saved to audio.mp3, then run Whisper on it:
whisper audio.mp3 --model turbo
  1. Read the result before you quote it. A 2024 study found that roughly 1% of Whisper transcriptions contained “entire hallucinated phrases or sentences.” The no-captions guide lists the signs to look for.

Downloading someone else's video raises YouTube's terms. The last question of that same guide sums them up.

Which Whisper model should you pick?

Whisper's README lists six sizes. Speeds are relative to the large model, measured on English speech:

The README also warns that Whisper's performance “varies widely depending on the language.” Run a two-minute clip in your video's language before you start a full lecture.

No graphics card? Run Whisper through an API

Groq hosts Whisper behind an API. Its documentation lists:

Cheap per hour. The cost is the setup: an account, an API key, and a script that sends the file and saves the text. On the free tier, audio over 25 MB has to be split or compressed first.

No terminal, no API key, and a video with no captions? That's the case I built Transcript Any Video for. It works straight from the YouTube link. You paste the link, and it transcribes the audio, not the captions. You get it word for word, with just your email, no password. Videos can run up to 2 hours, in 57 languages, detected on their own. You buy credits once, 10 for $1.99, no subscription, and you get a refund if the transcript is wrong.

Transcribe a video

What do the top results leave out?

Today, the results for this search are almost all tool pages, plus two video tutorials and a Maestra article last updated in December 2024. Of the three tool pages I could read, none says that YouTube's own transcript is already made by speech recognition, free to copy. None names the speech model it runs, so you can only compare them on promises. The Maestra article suggests Google Docs voice typing as a free method. Google's help describes voice typing through your microphone, and says nothing about audio your computer plays.

FAQ

Can Whisper translate a YouTube video into English?

Yes. OpenAI describes Whisper as a multitask model that does speech translation as well as speech recognition. Through Groq's API, only whisper-large-v3 translates. The turbo version there transcribes only.

Does Whisper label who is speaking?

No. Whisper writes one stream of text. A 2023 LessWrong comment found it “quite decent if there's only one speaker”, but noted it “doesn't do speaker diarisation or insert paragraph breaks.” For interviews and panels, some tools add speaker labels on top: YoutubeToText and ScreenApp both list them.

How long does it take to transcribe an hour of audio?

It depends on where Whisper runs. In 2023, Jeff Geerling found it “only a little slower than real-time” on his Mac Studio's processor, so about an hour. Groq lists a real-time speed factor of 189 for whisper-large-v3: on paper, under 20 seconds for an hour of audio, plus the upload.

Sources