Turn video into text

• Writing the caption for a Reel from what you actually said in it, instead of retyping it from memory.
• Pulling quotes from a live stream you recorded, a podcast clip or an interview to post as text or carousel slides.
• Checking the exact wording of a sponsor’s script, a giveaway’s rules or a price you mentioned on camera.
• Making notes from a long voice note or a recorded call, and finding the minute where something was said.
• Getting an SRT file for a video ad: Meta Ads Manager lets you add captions to a Facebook or Instagram video ad by uploading one.

Switch on timestamps when you need to find your way back into the recording: each line starts with the time it was said, such as [4:32]. Leave them off when you want clean paragraphs to paste into a caption, an email or a document.

Whisper is an open speech recognition model from OpenAI, released under the MIT licence. This page uses its two smallest multilingual versions: tiny (39 million parameters) as Fast and base (74 million) as Better. On a clear voice recorded close to the microphone, both get most words right: in our check with a 10-minute computer-generated English recording, Fast differed from the script on about 7 words in 100, mostly spelling variants such as color for colour, numbers and technical terms. Accuracy drops with music under the voice, several people talking at once, a phone held at arm’s length in a noisy street, and words the model rarely heard, like brand names and usernames.

Whisper knows 99 languages, and OpenAI’s model card says it performs unevenly across them, with lower accuracy where it had less training data. Auto-detect picks the language from the first 28 seconds with speech; if a recording mixes languages, or the detection is wrong, choose the language yourself and transcribe again. Hindi is where these small models struggled most in our check with computer-generated voices: Fast turned it into rough English, and Better managed a few words in Urdu and Latin letters before getting stuck. In Spanish, Better got all 28 words of a short sentence right and Fast got about a third wrong.

Tip: Numbers are the easiest thing to get wrong in a transcript. Before you post a price, a date or a discount code taken from one, play that moment back and check it.

• TXT is the transcript as you see it, with or without timestamps. SRT and VTT hold numbered subtitle lines of up to two rows of about 42 characters and no more than 6 seconds each.
• Files can be up to 20 minutes long, 600 MB for video and 100 MB for audio. The whole file is read into your browser’s memory to get at the sound, so on a phone keep videos to a few hundred MB.
• Speech recognition runs on your graphics chip through WebGPU where the browser offers it, and on the processor otherwise. If WebGPU gives output that can’t be speech, the tool redoes the job on the processor rather than show it.
• Your recording stays on your device. The speech model is downloaded from Hugging Face the first time and kept by your browser, so later transcripts start straight away.